fix(rate-limiter): exclude prompts from per-server tool quotas

[vishu-bh]

Enabling by_user_per_server currently rejects ordinary prompt fetches without server_id with RATE_LIMIT_CONTEXT_MISSING (503). This change confines catalog-server quotas to tool invocations and restores ordinary prompt retrieval.

Prompt evaluation uses a Rust engine configuration view with by_user_per_server disabled, sharing the existing backend and clock with tool evaluation. User, tenant, and tool/prompt-name counters remain shared across hooks. Tool identity validation, counter keys, algorithms, failure policies, and Python entry points are unchanged.

Adds memory and live Redis regression coverage for prompt quota exclusion and shared counters, updates hook-specific documentation, and bumps the package to 0.1.12 with matching Cargo lockfile and manifest versions.

Closes #202.

Validation

• make plugins-validate: passed (41 repository tests).
• Plugin make ci test-all test-integration: passed; 92 Rust tests and 134 Python hook tests, including live Redis and TLS, with no skips. Formatting, Clippy with -D warnings, release-wheel build/install, and benchmark compilation passed.
• Reproduction against installed CPEX: ordinary prompt without server_id succeeds; prompt and tool calls still share the global user quota.
• git diff --check: passed.

Local validation used nextest 0.9.133. The existing TLS fixture generates a certificate missing an Authority Key Identifier, which Python 3.13 rejects before calling the plugin. The final run used the fixture's supported RATE_LIMITER_TLS_REDIS_URL override with an isolated Redis endpoint and a valid temporary certificate; fixture code and certificate verification were unchanged. One existing one-second Redis TTL test failed once and passed on the full rerun.

Gateway REST GET/POST and MCP prompts/get acceptance tests remain pending in the gateway repository. This PR does not publish a release or update gateway dependencies.

[gandhipratik203]

LGTM!

[lucarlig]

LGTM