[vla, data] perf: pin pyav decoder to single thread

[XueSongTap]

Summary

Pin the PyAV video decoder to single-threaded decode (stream.thread_count = 1) in _decode_pyav, leaving all CPU parallelism to --num-workers.

PyAV defaults to thread_count=0 ("auto"), which sizes libav's thread pool to the host core count. Because the dataloader already provides process-level parallelism, this is a second parallelism knob layered over the same cores, and the two multiply: at 8 ranks x 16 workers on a 128-core host, 128 loader processes each spinning up ~128 decode threads oversubscribes the machine by orders of magnitude. The pool is also built and torn down on every av.open(), a cost that a handful of decoded frames never amortizes.

Thread-level decode is the worse knob for this access pattern in any case. Our workload is many small, independent seek-and-grab decodes; worker processes scale near-linearly across them, while libav slice threading yields ~1.6x at best on a single stream.

Impact

• +19% end-to-end training throughput
• Throughput curve is materially more stable. Previously, at --num-workers 16, throughput dipped once every 16 iterations as the dataloader round-robin came back around to a stalled worker; that periodic dip is gone.