Add prefill-stage concurrency limiter for PD mode

[james0zan]

Motivation

• Introduce a separate concurrency limiter for the prefill stage in PD (prefill+decode) routing to avoid prefill overload affecting the decode stage and to support queuing/timeouts for prefill requests.
• Expose the limiter via configuration and CLI so deployments can tune or disable prefill limiting independently from overall concurrency limiting.

Description

• Add max_prefill_concurrent_requests to RouterConfig with default -1, a builder setter, CLI flag --max-prefill-concurrent-requests, and validation in config/validation.rs to ensure the value is -1 or > 0.
• Add prefill_rate_limiter: Option<Arc<TokenBucket>> to AppContext and its builder with maybe_prefill_rate_limiter to initialize the limiter from config.
• Extend PDRouter with prefill_rate_limiter and queue_timeout_secs, implement a PrefillLimiterGuard helper that safely returns tokens on drop or explicit release, and integrate limiter acquisition/release around dual-dispatch logic.
• Change the dual dispatch send logic to use pinned futures and tokio::select! so the code can release the prefill token as soon as one of the requests completes (avoiding deadlocks/over-retention).
• Add unit tests for limiter guard behavior: test_prefill_limiter_guard_releases_on_drop, test_prefill_limiter_guard_release_now_idempotent, and test_prefill_limiter_guard_drop_unblocks_waiter, and update tests/fixtures to account for the new prefill_rate_limiter field on AppContext.

Testing

• Ran unit tests for the gateway package with cargo test -p sgl-model-gateway including the new pd_router tests; all tests passed.
• Existing PD router tests were verified to still pass after integrating the limiter and guard changes.

---

Codex Task