feat(server): export tracker metrics and scheduler queue telemetry

[Yumeio]

Expose /metrics as a projection of the same StatsTracker used by /v1/stats. Token counters are read directly from existing admitted/sampled totals; timing histograms observe each shared-generation RequestRecord once. A new scheduler -> tokenizer -> frontend queue snapshot supplies real running/waiting counts, including zero after abort/idle transitions. Unknown queue/pool values are omitted rather than fabricated, and registries are per tracker.

Timing preserves existing desktop behavior: streaming records have TTFT, buffered records do not. E2E covers shared chat/Messages/Responses generation; TPOT is the per-completed-stream mean after the first output. Raw /generate and legacy text completions contribute engine counters but are excluded from timing histograms because their middleware records before a stream finishes. See docs/server-hardening/03-metrics.md for semantics and vLLM/SGLang/TensorRT-LLM comparisons.

Validation on Windows x86-64 / Python 3.13.5: 568 passed for tests/server plus scheduler queue metrics, scheduler status and cost-accounting tests. All 5 tests/scheduler/test_abort_inflight_prefill.py tests also passed. New metrics/queue tests cover real Prometheus parsing, counter equality, late aborted tokens, ring eviction, repeated scrapes, IPC codecs and frontend telemetry. Targeted Ruff E4/E7/E9/F and git diff --check passed.

Reproduce: PYTHONPATH=python pytest -q tests/server tests/scheduler/test_queue_metrics.py tests/scheduler/test_scheduler_status.py tests/scheduler/test_cost_accounting_core.py tests/scheduler/test_abort_inflight_prefill.py. The Windows native CPU radix build required ninja on PATH and CL=/std:c++20; initial missing-dependency/compiler failures were resolved without changing kernel code. CPU/fake-backend tests only: no live GPU, checkpoint, multi-rank IPC or performance validation is claimed.

Adds prometheus-client>=0.20,<1 and additive IPC message types; run workers/frontend from the same checkout. No linked issue was supplied. AI-assisted draft at the user's request; human review and real-server validation remain required before merge.

Combined review evidence

All seven independent changes were also merged on the integration branch. The combined server and targeted scheduler run passed 779 tests; the branch records the configuration/error-handler conflict resolutions and detailed validation scope.

Validation host: Windows/Python 3.13.5 with existing PyTorch 2.13.0+cu126, outside the project's pinned 2.11 range. Results are scoped CPU server/state-machine evidence; they do not establish compatibility with the pinned production environment or real GPU/model execution.