fix(utils): tolerate coalesced msgpack frames in zmq pull queues

[Cyber-Marty]

Summary

ZmqPullQueue.get(), ZmqSubQueue.get() and ZmqAsyncPullQueue.get() decoded each received frame with msgpack.unpackb(frame, raw=False). When a frame carries more than one packed object, unpackb raises ExtraData — and both crash sites in #452 (freetoken/tokenizer/server.py:160 tokenize worker, freetoken/scheduler/io.py:85 scheduler) die on exactly that helper (utils/mp.py:68).

The reporter's hypothesis matches the failure profile: a coalesced read (two packed objects arriving in one frame) is non-deterministic, not tied to input, and the same helper serves both the tokenizer and scheduler receive paths. 2 crashes in 81 sequential requests.

Fix

Route every received frame through one buffered msgpack.Unpacker per queue (_CoalescedUnpacker):

• get() returns the first object from the buffer; a coalesced frame's remainder is held for the next call instead of crashing;
• empty() accounts for buffered objects (a socket-only check would report "empty" with undelivered messages pending);
• get_raw() / decode() stay unbuffered — the multi-rank broadcast path (_recv_msg_multi_rank0) pairs empty()/get_raw() with a rank-wide count, which a buffered remainder would desynchronize. Mixing them with buffered get() now raises a clear RuntimeError instead of silently skipping messages;
• the scheduler itself is unchanged — this only fixes the decode side.

Tests

New tests/utils/test_mp_coalesced.py (4 passed, 1 skipped):

• sync pull: a frame with two packed objects yields two get() results in order;
• sub queue: same through PUB/SUB;
• buffered-unpacker unit: feed once, take twice;
• empty() accounting with a buffered remainder;
• async variant skipped on Windows (zmq.asyncio needs a Selector loop; prod runs on Linux CI).

Red-verified: msgpack.unpackb on a coalesced frame raises ExtraData against the previous implementation.

Not in scope

get_raw()/decode() are deliberately left unbuffered with a mixing guard — the multi-rank broadcast path counts raw frames across ranks, and a buffered remainder would break that count. If coalescing is ever observed on that path, it needs a design pass (length-prefix framing) rather than a local patch.

Fixes #452