[python] Configure LeRobot video decoder cache through table options

[XiaoHongbo-Hope]

Random sampling across more than eight videos can repeatedly evict and reopen decoders. Increase the default cache from 8 to 16 per camera per worker, and expose read.video.max-open-decoders as a dynamic table option. Use table.copy({"read.video.max-open-decoders": "8"}) to trade decoder reuse for a smaller cache without changing stored table options.

On the public lerobot/koch_pick_place_5_lego dataset, a paired run changing only decoder capacity improved sustained throughput from 69.8 to 90.8 samples/s (+30.1%). Both runs used CPU TorchCodec decoding, OSSFS video reads, Rust table reads, a 256 MiB worker-local block cache, batch size 32, four workers, and eight-frame uint8 windows; no image transforms or model compute. Each run warmed up 10 batches and measured 6,400 samples. Sample order and first/last batch tensors matched. This is one paired measurement, not a cold-start result; larger decoder caches can retain more memory.

Tests: 164 passed, 50 skipped across the LeRobot, multimodal table, and video modules; option-name regression tests, Flake8, and git diff --check passed.