Reading poses in parallel multithreaded may cause errors

[AmitMY]

Followup from sign-language-processing/spoken-to-signed-translation#46

It seems like reading poses in parallel may result in errors.

https://github.com/sign-language-processing/spoken-to-signed-translation/blob/21a34fbb7ae6439eb8ed54b0c4a2a5c4538a7977/spoken_to_signed/gloss_to_pose/lookup/lookup.py#L111-L130

We should add a test

[AmitMY]

Fixed in #239: the root cause was PoseHeaderCache being updated non-atomically — concurrent Pose.read calls on files with different headers could crash ("buffer is too small for requested array") or silently return a pose parsed with another file's header. All cache access now goes through a lock, and check_cache returns the header and end offset from the same critical section. The test this issue asked for exists as test_concurrent_read_of_different_files_is_safe (fails in under a second on the old code; an external repro went from 122 exceptions + 153 wrong headers to 0 across 4,800 concurrent reads).