…ule checkout

beta.7's release run surfaced three failures, all consequences of the
vendoring landing without a real compile/CI pass (CI only runs on `main`,
so the branch never exercised them). This fixes all three; the version stays
at 0.6.0-beta.7 so the existing tag/release can be deleted and recreated.

1) CI Engine job (clippy) — the real blocker
   `cargo clippy -- -D warnings` linted the vendored llama-cpp-rs crates
   because a path dependency living under the workspace root is ALWAYS a
   workspace member (verified: cargo `exclude` and a nested `[workspace]`
   do NOT demote it), and RUSTC_WORKSPACE_WRAPPER runs clippy on every
   member. Upstream's build.rs then tripped clippy::uninlined_format_args
   and missing_docs, which `-D warnings` turned into hard errors.
   Fix: `cargo clippy --no-deps` (lint only our own crate). Proven locally
   — full CI-equivalent clippy now exits 0; the vendored crate's own
   warnings are shown but no longer gate. Belt-and-suspenders: the vendored
   manifests are now self-contained (dependency versions inlined from
   upstream's [workspace.dependencies]) and drop `[lints] workspace = true`,
   so even when compiled as members they carry no pedantic/missing-docs
   opt-in. The root workspace.dependencies/workspace.lints tables (added in
   beta.7) are removed as no longer needed. Dangling upstream `[[example]]`
   targets (pointing at a non-vendored examples/ dir) are dropped too.

2) build-cuda (Linux CUDA, container) — submodule checkout
   nvidia/cuda images ship without git, so actions/checkout fell back to the
   REST API tarball, which cannot fetch submodules ("Input 'submodules' not
   supported when falling back to download using the GitHub REST API"). Added
   a git-install step BEFORE checkout so it does a real clone. This is a
   container-only issue: the standard Linux/macOS/Windows runners have git and
   their builds passed (proving the vendored crates + new llama.cpp pin compile
   cross-platform).

3) build-windows-cuda — transient cache-service failure
   The job died at Swatinem/rust-cache restore (GitHub Actions cache backend
   blip), before any compile, nuking the ~50 min long-pole. Added
   `continue-on-error: true` so a cache outage degrades to a slower cacheless
   build (sccache/S3 still carries the C++/CUDA objects) instead of failing.