metal: fix the async-output completion-event wedge

[auroter]

Fixes the Metal serving bug where the engine could hang forever on a boot's first large prefill request, which until now required a boot-ramp workaround (serve a short request with real decode steps before any multi-chunk prefill, or the boot was lost and needed a restart).

Stacked on #2 — this branch contains the M1 Ultra campaign commits until that PR merges; the fix itself is the single top commit (async_utils.py + metal_compat.py docs + notebook).

The bug

On timing-sensitive boots the engine parked forever inside MPSEvent::synchronize waiting on the async-output copy event: GPU idle, no command-buffer errors, the completion signal simply never arrived. Reproduced deterministically in the 08-14 bisect (2/2 boots) and again during the 08-15 merge gate. Once parked, the boot was unrecoverable.

Root cause

Apple exposes one GPU queue, and torch.Stream(mps) always returns the same stream (id 0). The shared async-output path still ran the full CUDA cross-stream choreography — switch to a "copy stream", wait_stream on the "main stream", record a generic torch.Event on the copy stream. On Metal that dance buys zero overlap (there is only one stream) and routes every step's completion through an event that can lose its signal.

The fix

On Metal, skip the choreography entirely: the same non-blocking device-to-host copies enqueue on the one real stream, and completion is a native torch.mps.Event recorded there — no stream switching, no cross-stream waits. The CUDA/ROCm path is byte-identical. An ops fallback (VLLM_QC_ASYNC_OUT_DRAIN=1) replaces the event with a full stream drain; it costs about 4% decode throughput, so it is not the default.

Validation (dsv4-xxs-1, M1 Ultra, five fresh boots)

• All pinned anchors bit-exact: 8-tok, 1k-in/2k-out, and the 2500-token multi-chunk anchor.
• Decode at 32.0 tok/s end-to-end on the 2k-token anchor — the best wall of the day, no regression from the fix (the drain-only variant measured 30.6 tok/s, which is why it is only the fallback).
• Both historical wedge-trigger protocols run clean on the fixed build: cold boot straight to a multi-chunk request, and primer + short request then multi-chunk.
• Kill-switch boot serves bit-exact.

One honest caveat: the wedge is a host-timing race and had drifted out of its trigger window on the test box this week (the unfixed build also ran the triggers clean today). The fix rests on the structural argument — after it, there is no cross-stream event on the completion path left to lose — plus the earlier deterministic bisect evidence. The boot-ramp ops protocol stays recommended until a longer soak confirms it is no longer needed.