Add async request queue to llm-server

[stikves]

Summary

The llm-server currently gates concurrency with a binary busy flag: while one
request is generating, any concurrent request is rejected immediately with
429. This PR replaces that gate with an async RequestQueue, so concurrent
requests wait in a bounded queue and are rejected only when the queue is full.

Changes

• RequestQueue (CoreAILMCommon): an async semaphore built on
  withCheckedThrowingContinuation with an acquired / rejected / queued state
  machine. Replaces tryAcquire()/release() and the generating flag.
  acquire() returns a QueuePermit that frees the slot exactly once — on
  explicit release() or, as a backstop, on deallocation, so a dropped
  streaming response can never leak the slot and wedge the server.
• Queued waiters are cancellation-aware: if the awaiting task is cancelled
  while queued (e.g. on shutdown), the waiter is removed and its acquire()
  throws CancellationError instead of leaking its continuation or being handed
  the slot after the fact.
• --max-queue-depth (default 16): how many requests may wait before the
  server returns 429 with a queue_full error. Rejected as invalid if
  negative; 0 means no queuing (second concurrent request gets an immediate
  429).
• hasRecurrentState on InferenceEngine (default false; overridden by
  the sequential and pipelined engines): recurrent (SSM / hybrid) models can't
  reuse a token prefix, so the server skips the prefix-reuse fast path and counts
  a miss. Engine correctness on rewind is already enforced by the engine itself;
  this only keeps the server's reuse stats and logs accurate.
• ServerError.queueFull(depth:) for the rejection path.

Behavior

Scenario	Before	After
Single request	Served	Served (unchanged)
Concurrent requests	Immediate 429	Queued, served FIFO
Queue at capacity	Immediate 429	429 (queue_full)

/ready now reports busy based on queue activity rather than the removed
generating flag.

Testing

• RequestQueue unit coverage (in CoreAILMCommonTests): immediate acquire,
  concurrent waiting, FIFO wakeup, rejection at maxDepth, maxDepth == 0,
  negative-depth clamp, idempotent permit release, and cancellation of a queued
  waiter. CoreAILMCommonTests (52) pass; llm-server builds.
• Load-tested end to end against muse_glimmer_30b_4bit_dynamic:
  • Single request: served.
  • 8 concurrent (< depth): all served FIFO, no premature 429.
  • 30 concurrent (> capacity): 17 served, 13 rejected with queue_full. This
    matches 1 in-flight + 16 queued at the default depth.
  • 20-request soak: no errors; server returned to idle with no stuck slots.