perf(ci): share one bounded Rust cache

[Teakowa]

What changed

• add one shared linux-ci Rust cache family for LPP CI;
• make the rust job the only cache writer, and only on main;
• keep PRs and the conformance job restore-only so cache growth stays bounded;
• use the committed Cargo.lock consistently with --locked;
• run conformance binaries from the normal debug profile instead of compiling a separate release profile solely for protocol validation.

Why

The current rust and conformance jobs run in parallel on separate fresh runners and independently download/compile the same workspace baseline. The previous artifact-based attempt removed the duplicate build but serialized the jobs, which improved runner work at the cost of CI latency. This revision keeps the jobs parallel and uses a single bounded cache instead.

There is no release-artifact or performance contract in this workflow: conformance validates protocol behavior against fixtures and the mock provider, so a dedicated optimized build did not provide additional coverage.

Cache policy

• one logical cache family: linux-ci;
• rust saves only on refs/heads/main;
• conformance never saves;
• failed jobs do not save;
• PRs restore but never create long-lived cache entries.

Validation

• conformance remains a separate parallel job;
• fixture paths and provider invocation are unchanged;
• CI on the latest commit validates the shared-cache configuration and debug-profile conformance behavior.

Fixes #7