Add one-shot CLI over shared rust docs runtime

[snowmead]

Summary

• Add a transport-free RustDocsRuntime shared by the MCP service and CLI paths.
• Add a one-shot CLI adapter via rust-docs-mcp call <tool> --params <json>.
• Keep default/serve behavior as the existing long-running MCP stdio server.
• Add blocking CLI cache behavior so call cache-crate completes download, rustdoc generation, and search index creation before exiting.
• Add runtime/CLI integration tests and README examples.

CLI examples

rust-docs-mcp serve
rust-docs-mcp call cache-crate --params '{"crate_name":"semver","source_type":"cratesio","version":"1.0.0"}'
rust-docs-mcp call search-items-fuzzy --params '{"crate_name":"semver","version":"1.0.0","query":"Version","limit":5}'

Verification

• cargo fmt -p rust-docs-mcp -- --check
• cargo build -p rust-docs-mcp
• cargo clippy -p rust-docs-mcp
• cargo test -p rust-docs-mcp --no-run
• cargo test -p rust-docs-mcp --test integration_tests test_runtime_cache_blocking -- --nocapture
• cargo test -p rust-docs-mcp --test integration_tests test_runtime_search_fuzzy -- --nocapture
• cargo test -p rust-docs-mcp --test integration_tests test_cli_call -- --nocapture
• End-to-end binary smoke test with fresh temp cache:
  • call cache-crate
  • call search-items-fuzzy
  • call list-cached-crates
• Full integration suite:
  • Default parallel run hit the existing 30s timeout in test_cache_update under resource contention.
  • Reran test_cache_update alone: passed.
  • Reran full integration suite serially: 27 passed; 0 failed.

Notes

cache_operations remains MCP-oriented because task state is in-memory and only meaningful for a long-running process. The CLI cache-crate path is intentionally blocking.