feat: record FastMCP HTTP request and tool call metrics

[lesnik512]

Builds on #280's _postprocess_http_apps hook (merged).

Why

FastMcpPrometheusInstrument only serves the registry, so a FastMCP service's /metrics has nothing but process_* and python_*. FastAPI, Litestar and FastStream all record request metrics. Ported from community-of-python/microbootstrap#155.

Changes

• HTTP requests. Every ASGI application http_app() builds is instrumented with prometheus-fastapi-instrumentator, the library the FastAPI bootstrapper already uses (it needs only Starlette). Requests are labelled by route template (handler="/mcp", handler="/health/"), unknown paths are grouped into handler="none", and the metrics route is excluded by default. prometheus_instrumentator_params and prometheus_instrument_params are passed through under the same names as in FastAPIConfig.
• Tool calls. FastMcpPrometheusMiddleware records fastmcp_tool_calls_total{tool, status} (success / error) and fastmcp_tool_call_duration_seconds{tool}. It is on by default, and fastmcp_prometheus_tool_metrics_enabled=False turns it off.
• The fastmcp-metrics extra adds prometheus-fastapi-instrumentator>=8.0.1, the same floor as fastapi-metrics. The instrument now needs both packages; with either missing it is skipped with the usual dependency warning.
• Docs: a "Prometheus FastMCP" configuration section and a metrics section in the FastMCP guide.

Registry notes

• When http_requests_total and its siblings are already registered, metrics.default() does not reuse them. It catches the duplicate and returns None, so a second instrumented application records nothing. The instrument therefore builds the default metrics once, on the first http_app(), and .add()s that closure to every later application. Its parameters are the subset of instrumentator and instrument params that metrics.default accepts.
• Across bootstrappers in one process, the global-registry limit is the same one FastAPI already has. The test module drops the instrumentator's collectors before each test for this reason.
• The tool-call collectors reuse already-registered collectors by name, so a later bootstrap in the same process (or a module reload in tests) does not raise DuplicateTimeseries.

Tests

• Requests are counted by route template, none is counted for unmatched paths, and the metrics route is not counted.
• Two http_app() calls are both counted. This test fails without the shared default metrics.
• prometheus_instrumentator_params reach the instrumentator.
• Successful and failing tool calls and their durations are counted, and tool metrics can be turned off.
• Teardown restores http_app. The instrument is skipped without the instrumentator.
• just lint, ty 0.0.85 and just test-ci pass on top of feat!: trace Litestar through its own OpenTelemetryPlugin #279 and feat: trace FastMCP HTTP requests with OpenTelemetry #280 (358 passed, 100% coverage).
• Floor smoke passes (fastmcp==3.0.0, prometheus-client==0.20.0, prometheus-fastapi-instrumentator==8.0.1, Python 3.11). At those floors, two http_app() applications each serving the health route and a 404 count 2 and 2, and a tool call counts 1.