Add native Prometheus serving metrics endpoint

[remer]

Adds a non-queued GET /metrics endpoint to the Rust HTTP server so telemetry remains scrapeable during long inference.

Exposed signals:

• exact model-worker running and scheduler waiting gauges
• prompt/generation token counters
• exact prefix-query and reused-prefix token counters
• active logical KV tokens and context occupancy
• TTFT, inter-token latency, and end-to-end latency histograms
• terminal outcomes (stop, length, tool_calls, error, abort)
• RAM session-cache occupancy, hit/miss/eviction, and save/restore counters

The common serving metrics use the established vllm: names for collector compatibility; Krasis-specific occupancy/session metrics use the krasis_ namespace. The endpoint snapshots metrics under a short mutex and formats after releasing it. Per-token TPOT and logical-KV updates use one lock acquisition.

Semantics:

• TTFT starts at HTTP admission and includes scheduler wait.
• TPOT is the interval between consecutive generated tokens.
• logical KV occupancy counts tokens actually represented in active sequence state; it is intentionally not the percentage of the preallocated KV buffer.
• prefix hits are a token numerator and prefix queries are full prompt-token denominator.
• a request guard records early exits as errors and clears running/KV gauges on every path.

Validation:

• cargo check --release --lib passes with CUDA enabled
• optimized extension build imports under Python 3.13
• live smoke covered non-streaming completion, streamed length completion, exact progressive-prefix reuse, terminal counters, histogram counts, and a 10 ms /metrics scrape while generation was active
• no credentials, hostnames, addresses, local paths, or request contents are included in this change