Adds the `PackageImportMethod` variant to `LogEvent` and emits it from
`Install::run` after `pnpm:context`, before `pnpm:stage importing_started`.
Mirrors pnpm's emit at
https://github.com/pnpm/pnpm/blob/086c5e91e8/fs/indexed-pkg-importer/src/index.ts#L32.

The wire-format enum has only `clone | hardlink | copy`; pacquet's
`Npmrc::package_import_method` adds `Auto` and `CloneOrCopy`. Both
collapse to `clone` because that's the optimistic path the upstream
auto-importer attempts first. For the explicit `Hardlink` / `Copy` /
`Clone` settings the wire value matches reality. A TODO at the emit
site notes that the value should be re-derived after the first
successful import once the fallback signal is surfaced past the
per-package import path.

While auditing this channel, also deferred `pnpm:stage
resolution_started/done`: pacquet's no-lockfile flow interleaves
resolution and import per dependency, so emitting `resolution_done`
after import work has already happened would break the JS reporter's
per-package state. Pick this back up once pacquet's no-lockfile flow
gains a separate resolution pass. Frozen-lockfile installs (pacquet's
primary path today) don't need these events anyway — pnpm itself skips
them in `deps-restorer`. The deferral is recorded under #347.

Tests:
- New `package_import_method_event_matches_pnpm_wire_shape` in the
  reporter crate asserts every wire value renders correctly.
- The install integration test grows to assert the five-event success
  sequence (Context, PackageImportMethod, ImportingStarted,
  ImportingDone, Summary) and spot-checks the wire mapping for the
  default `Auto` config (collapses to `Clone`).
- Verified the test catches the regression by temporarily commenting
  out the new emit and confirming the failure shape, then restored.

Refs #347. Engine: #345. Previous channel: #355.

Per review feedback on #356: the doc comment on
`package_import_method_event_matches_pnpm_wire_shape` cited
`cloneOrCopy` (camelCase) as the spelling pacquet's config enum
deserializes from. `pacquet_npmrc::PackageImportMethod` is annotated
with `rename_all = "kebab-case"`, so the on-disk and config spelling
is `clone-or-copy`. Update the example to match what `.npmrc` actually
contains.

Drop the redundant `"method {expected}"` failure message from the
inner `assert_eq!`; `assert_eq!` already prints both sides and the
extra context wasn't worth keeping.

Per coderabbit review on #356: the previous commit kept
`LOGGED_METHODS` as a module-static, which suppresses the
`pnpm:package-import-method` reporter event for every install after
the first in the same process. That contradicts upstream pnpm, where
`createIndexedPackageImporter` builds a fresh closure per install, so
the dedupe state is install-scoped.

Threads an `&AtomicU8` from `Install::run` down through
`InstallFrozenLockfile`, `InstallWithoutLockfile`, `CreateVirtualStore`,
`CreateVirtualDirBySnapshot`, `InstallPackageBySnapshot`,
`InstallPackageFromRegistry`, `create_cas_files`, and
`link_file`. Each install layer carries the atomic as a struct field
(matching the existing `verified_files_cache` pattern). The atomic
type is `Sync`, so the rayon par_iter and tokio futures in the chain
work without further changes. Removes the module-static
`LOGGED_METHODS`.

The unit test on `log_method_once` already passed its own atomic, so
its contract is unchanged. Other `link_file` tests that don't care
about emit semantics now construct an inline `&AtomicU8::new(0)` as a
throwaway dedupe state; they exercise `SilentReporter` so the atomic
is never read for emission anyway.

Practical effect today is zero — pacquet runs as a one-shot CLI and
exits after one install — but the change removes a footgun for tests
that exercise multiple installs in the same nextest binary, and for
any future embedded use.

Per Copilot review on #356:

- The `LOG_FLAG_*` constants in `link_file` were widened to
  `pub(crate)` in the previous commit but they're only referenced
  from within this module (the inline `mod tests` can already see
  parent private items). Restore them to module-private to keep the
  crate's internal API surface tight.

- `Reporter::emit` may be invoked concurrently from arbitrary
  threads — `link_file` is called from `create_cas_files`'s rayon
  `par_iter`, and tarball / store-index work runs across tokio
  workers. Document that contract on the trait so future reporter
  implementations don't reach for unsynchronized global state.
  Note that both production sinks (`SilentReporter`,
  `NdjsonReporter`) already satisfy it.