Start CPU workers safely and pass them state per call

[joeljose]

Closes #69

Changes

• Start method: workers use forkserver (spawn where it doesn't exist), never plain fork. The fork server preloads numpy, scipy.fft/ndimage, dtcwt and cv2, so workers share those pages. The CPU path now also parallelises on platforms without fork; the old code fell back to serial there.

• No module state: _POOL is gone. Each magnify_motions call allocates its input, lowpass, per-level highpass and result arrays with _SharedArrays:

  • Linux: anonymous in-memory files (memfd_create), which workers open via /proc/<pid>/fd/<n>.
  • Elsewhere: temp files.

  Workers receive only paths and memory-map them. jobs=1 uses plain arrays and no pool.

• Why memfd and not /dev/shm or multiprocessing.shared_memory: Docker's default /dev/shm is 64 MB. My first attempt with files there crashed workers (BrokenProcessPool) on real clips. memfd isn't limited by that mount.

• Pool lifetime: one pool per call, used for both the forward and inverse passes. The result is returned as an ndarray view of its mapping, with no copy. File descriptors are released when the call ends, and the result's mapping when it is freed.

• _transform uses functools.lru_cache (per process).

• README notes that scripts calling magnify_motions with jobs > 1 need an if __name__ == "__main__": guard (standard for forkserver/spawn). Performance numbers and CHANGELOG updated.

Measurements (face.mp4, default settings)

	before (fork)	after (forkserver)
--jobs 2 time	55.5–56.2 s	59.5–60.2 s (+7%)
--jobs 2 peak (cgroup)	2.32 GiB	2.49 GiB (+7%)
--jobs 1 time	87.9 s	84.4 s
output, --jobs 1 vs --jobs 2	identical	identical (0 of 282M values differ)

The cost comes from each worker's own Python heap and the fork server process. estimate_memory still matches (2.48 GiB estimated).

Tests

• New: two concurrent calls from threads each return their own correct result (vs serial); the start method is not fork; memfds are released (only the returned result's mapping remains, and it goes when the result is freed).
• CPU image: 113 passed, 3 skipped. GPU image on an RTX 4050: 137 passed, 1 skipped.