Adds the zip-archive side of pnpm's binary fetcher. Mirrors
[`downloadAndUnpackZip` and `extractZipToTarget`](https://github.com/pnpm/pnpm/blob/94240bc046/fetching/binary-fetcher/src/index.ts):
walks zip entries, rejects absolute / `..` paths via
`zip::ZipFile::enclosed_name`, strips the archive's top-level
basename (`BinaryResolution::prefix`) so the ignore filter and CAS
keys see the same relative paths upstream's regex does, then
writes each file entry into the CAS through `StoreDir::write_cas_file`.

* `extract_zip_entries` — parallel to `extract_tarball_entries`,
  with `archive_prefix` + `ignore_file_pattern` plumbing
* `fetch_and_extract_zip_once` / `fetch_and_extract_zip_with_retry`
  — same network permit shape, retry policy, and post-download
  semaphore gating as the tarball path
* `DownloadZipArchiveToStore` — sibling of `DownloadTarballToStore`
  with the same store-index lookup, prefetch cache reuse, and
  store-index writer queue
* `ignore_file_pattern` threaded through `DownloadTarballToStore`
  and the tarball extractor (the tarball half of upstream's
  `archiveFilters` map)
* Adds `zip = "5", default-features = false, features = ["deflate"]`
  to the workspace; reused in `pacquet-tarball`

Includes 4 unit tests for the zip extractor: prefix-stripping,
ignore-filter applied to the stripped path, `..` rejection, and
the no-prefix passthrough. Each rejection guard was verified by
temporarily disabling it and re-running the test.

The install-dispatch site (slice D) and the per-package
`ignoreFilePattern` value (the Node-runtime extras filter) will
land in the follow-up PR.

Address Copilot review on #472:

- `extract_zip_entries` masks `entry.unix_mode().unwrap_or(0o644) & 0o777`
  before writing `CafsFileInfo.mode`, matching the permission-only
  convention enforced by `store_dir::add_files_from_dir::file_mode_from`.
  Without the mask, the high `st_mode` bits (e.g. `0o100000` for a
  regular file) would land in the `index.db` row and diverge from
  what pnpm writes for tarball entries / on-disk imports.
- Adds `TarballError::ReadZipEntries(std::io::Error)` for zip
  per-entry I/O failures (try_reserve / read_to_end), distinct from
  `ReadTarballEntries`. The retry-classification path now emits
  `ERR_PACQUET_ZIP` rather than the misleading
  `ERR_PACQUET_TARBALL_TAR` for zip-side failures.
- Changes `ignore_file_pattern` from `Option<&'static IgnoreEntryFilter>`
  to `Option<Arc<IgnoreEntryFilter>>` on both `DownloadTarballToStore`
  and `DownloadZipArchiveToStore`. The Slice D dispatcher can now
  construct one filter per fetch from runtime config (the per-package
  `archiveFilters` map upstream uses) without leaking the closure or
  pinning it to `'static`. Cloning the Arc per retry attempt is cheap;
  the inner trait object is shared.

`extract_zip_entries` ran the `is_dir()` early-continue *before*
`enclosed_name()` validation, so an archive carrying a directory
entry like `../evil/` was silently skipped instead of surfacing
`TarballError::PathTraversal`. Pacquet wouldn't have written that
directory either way (only file entries take the CAS write path),
but the contract is "reject any unsafe entry" — for tooling that
inspects the error code.

Move the path validation above the `is_dir()` check. Adds a unit
test `extract_zip_rejects_directory_entry_with_parent_component`
that catches the regression (verified by temporarily reverting the
guard order).

Caught by CodeRabbit on #472.

`[crate::BinaryResolution::prefix]` resolved against `pacquet_tarball`
where `BinaryResolution` doesn't live — it's `pacquet_lockfile::BinaryResolution`.
Switched both references to text + fully-qualified path so rustdoc
under `-D warnings` (the CI Doc job) stops failing.

…nostic OOM error

Address Copilot review on #472:

- `extract_zip_entries` builds the canonical `cas_paths` /
  `pkg_files_idx` key from `enclosed_name()` rather than the raw
  `entry.name()`. `enclosed_name()` collapses `.` segments and is
  already path-traversal-checked, so the map key, the ignore-filter
  input, and the basename strip all see the same string.
  Cross-platform: rebuild from `Normal` components joined with `/`
  so a Windows host can't smuggle `\` separators into the key.
- `TarballError::TarballTooLarge` display text changed from
  "Tarball at {url} ..." to "Archive at {url} ..." so the zip
  pipeline (which reuses `allocate_tarball_buffer`) doesn't surface
  a misleading "Tarball ..." message on OOM / oversize. Diagnostic
  code stays `pacquet_tarball::tarball_too_large` for variant
  identity.
- `run_with_mem_cache` doc-blocks the "stable filter per URL"
  invariant: the mem cache keys on `package_url` alone (matching
  pnpm's `tarballCache`), so callers must keep the (URL, filter)
  relation functional. Tarball URLs encode (name, version,
  integrity) and `archiveFilters` is keyed by `pkg.name` upstream,
  so this naturally holds; the Slice D dispatcher follows the same
  table.
- New unit tests:
  - `extract_tarball_applies_ignore_filter_dropping_entries_from_both_maps`
    covers the tarball-side filter path (the existing zip tests
    only exercised it through `extract_zip_entries`).
  - `extract_zip_normalizes_dot_segments_in_entry_paths` pins the
    `.`-segment collapse. Verified by temporarily falling back to
    `raw_name`.

Address Copilot review on #472:

- `fetch_and_extract_zip_once` was hitting `client.get(package_url)`
  without resolving `AuthHeaders::for_url`, so a runtime archive
  hosted behind a token-protected proxy would 401 even when
  credentials were configured. Threaded `auth_headers:
  &AuthHeaders` through `DownloadZipArchiveToStore`,
  `fetch_and_extract_zip_with_retry`, and `fetch_and_extract_zip_once`,
  and attach the `authorization` header the same way the tarball
  path does.
- `TarballError::ReadZipEntries` only wrapped `std::io::Error`,
  so the rendered error said "Failed to read zip entry: <io_err>"
  with no URL or entry path. The accompanying doc claims pacquet
  surfaces the entry path that triggered the failure (matching
  `ReadZipArchive` and `PathTraversal`), so promote the variant
  to a struct shape carrying `url` + `entry_path` + `source` and
  attach the cleaned path at both call sites (the `try_reserve`
  for the entry's payload and the `read_to_end`).