Correctness, robustness, and quality improvements

[dahankzter]

Summary

A batch of fixes, performance improvements, and quality-of-life changes found during a systematic audit of the codebase. Each commit is self-contained and independently reviewable.

Correctness fixes

• Replace unwrap with proper error on empty gRPC response — a peer returning GetResponse { value: None } previously caused a panic; now returns a descriptive error
• Add default max capacity to main_cache — without a bound, the main cache could grow without limit and exhaust memory (default: 100k entries)
• Store SocketAddr directly in VNode — eliminates a parse().unwrap() that could panic on malformed strings; custom Hash impl preserves existing ring distribution
• Eliminate TOCTOU race in add_peer / remove_peer — double-checked locking prevents duplicate peer entries under concurrent service discovery
• Don't hold Arc across sleep in service discovery loop — the previous while let Some(cache) = cache.upgrade() pattern kept a strong reference alive during sleep, preventing Drop from firing;
  restructured to loop { sleep; upgrade }

API improvements

• Make GroupcacheError a public enum — consumers can now pattern-match on error variants (Loading, Transport, Deserialization, etc.) instead of relying on string matching
• Implement Display for GroupcachePeer — useful for logging and diagnostics
• Log remote load errors — remote load failures that fall back to local load were previously silent; now logged at warn level

Performance

• Use Arc<str> instead of String for cache keys — avoids heap allocation on every cache lookup/insert; reference-count increment replaces full string clone
• Use rmp_serde::from_slice instead of from_read — zero-copy deserialization path for byte slices

Housekeeping

• Abort service discovery task on shutdown — GroupcacheInner now implements Drop to abort the background service discovery task via OnceLock<AbortHandle>
• Clean up build.rs — remove tonic-prost-build from runtime [dependencies] (kept only in [build-dependencies])
• Add Makefile — targets for build, test, lint, coverage (lcov/html/summary), clean, and help
• Comprehensive test coverage — 15 new unit + integration tests bringing line coverage to ~98%; generated proto code excluded from reports

Test plan

• All 46 tests pass (cargo test --workspace)
• Clippy clean (cargo clippy --workspace -- -D warnings)
• Formatting clean (cargo fmt --all -- --check)
• ~98% line coverage on library code (cargo llvm-cov --workspace --all-features --ignore-filename-regex 'groupcache_pb\.rs')
• Verify no behavioral regressions in downstream consumers