Fix websocket emit-queue starvation when emit() returns None

[dr-regier]

Problem

WebSocketHandler._emit_to_queue enqueues whatever emit() returns with no guard:

while not self.quit.is_set():
    output = await self.stream_handler.emit()
    self.queue.put_nowait(output)

When a handler is idle, emit() returns None instantly on every call. With an unbounded queue and no yield on the idle path, this spins thousands of times/sec enqueuing None, floods the queue (observed ~164k depth), and starves the consumer _emit_loop so it isn't scheduled to send the frame already at the head of the queue.

Real-world impact (telephony): up to ~18s of intermittent dead air before an already-generated greeting reached the caller. Related: #203, where the same starvation is worked around by adding await asyncio.sleep(0.01) inside the user's emit().

Fix

On the idle path (output is None), back off 20ms and skip the enqueue:

if output is None:
    await asyncio.sleep(0.02)
    continue
self.queue.put_nowait(output)

Why this is safe

_emit_loop already skips None outputs, so enqueuing None never produced a frame - it only created work and event-loop contention. This removes wasted work and yields the loop to the consumer on idle; no downstream behavior changes.

Validation

On the live phone path, first-frame send dropped from ~18.6s to ~0.1-0.4s with queue depth ~0 on every call. Tested on 0.0.33; the affected code is identical on main.

Notes / alternatives

• 20ms is a conservative idle backoff (negligible added latency vs the starvation it prevents). Open to tuning.
• Alternatives considered: bounding the queue, or documenting the emit() await requirement instead of handling it here. Went with the loop-side guard because it's minimal and fixes the existing AsyncStreamHandler breaks when no await is used in emit() #203 workaround at the source. Happy to reshape.