Fix PoseHeaderCache race under concurrent Pose.read

[AmitMY]

Summary

PoseHeaderCache is global mutable state updated non-atomically, which makes Pose.read unsafe to call from multiple threads when the files have different headers:

• PoseHeader.read re-read PoseHeaderCache.end_offset after check_cache returned — a concurrent set_cache in between leaves the reader at another file's body offset.
• set_cache assigned header/offsets before computing hash, so a concurrent check_cache could match the stale hash and return the newly-assigned wrong header.

Downstream this is the intermittent TypeError: buffer is too small for requested array seen in spoken-to-signed-translation CI (its lookup_sequence reads lexicon poses in a ThreadPoolExecutor; see e.g. sign-language-processing/spoken-to-signed-translation#52 discussion). The nastier failure mode is silent: a pose parsed with a different file's header and no exception.

Repro (6 lexicon files, 8 threads, 4,800 reads): 122 exceptions + 153 silently-wrong headers before the fix; 0 + 0 after.

Fix

The authoritative cache is now a single immutable (hash, header, start_offset, end_offset) tuple swapped atomically (CPython attribute assignment); readers take one snapshot and use only it. No locks needed. The legacy class attributes are kept in sync for backward compatibility (Pose.read uses end_offset as a read-ahead hint, where a torn read is harmless).

Test plan

• New regression test test_concurrent_read_of_different_files_is_safe (fails in <1s on the old code, passes with the fix)
• pytest tests/pose_test.py pose_format/utils/reader_test.py — all pass (except pre-existing test_unpack_tensorflow, which needs tensorflow installed)
• External repro script against this branch: 0 failures in 4,800 concurrent reads

🤖 Generated with Claude Code

[AmitMY]

Reworked to a plain threading.Lock as suggested (stdlib, so no new wrapt dependency). One subtlety: the lock can't only guard set_cache/clear_cache — the reader is the other half of the race. check_cache reads hash/start_offset/end_offset/header as four separate attribute reads, and PoseHeader.read then re-read end_offset after the check, so a locked writer can still be observed mid-update by an unlocked reader. So check_cache takes the same lock and returns (header, end_offset) from inside the critical section. Re-verified: the concurrency regression test passes and the external repro is clean (0 failures in 4,800 concurrent reads; 122 exceptions + 153 wrong headers before the fix).