Skip short trailing windows so batch > 1 collate sees uniform shapes

[jeqcho]

Fixes #80.

Root cause of the trailing-batch crash: near an episode tail, get_frame_indices trims the window (unique_sorted = unique_sorted[:-7]), producing a sample with fewer chunks than the rest of the batch. With per_device_train_batch_size > 1 the np.stack in dreamzero_cotrain.py then sees mixed shapes and raises ValueError: all input arrays must have the same shape. On our corpus this fired roughly once per 580 batches, so short smoke runs pass and long runs die.

Fix: skip short windows entirely (same mechanism as the existing <= 8 frames case) instead of trimming to a different chunk count. The empty array needs an integer dtype because it is used as an index downstream. The cost is discarding a small number of episode-tail anchors, which seems preferable to a crash; padding them to full length would change training semantics.

Note this is distinct from the variable-chunk design of the synthetic DROID dataset discussed in #69 — this path is the sharded LeRobot loader, where every sample is supposed to be full-length.

Validated on a 12k-step fine-tune of DreamZero-AgiBot at per-device batch 4 (global 32) over 124 bimanual YAM LeRobot repos with zero collate failures: https://github.com/robocurve/dreamzero-yam, released at https://huggingface.co/robocurve/dreamzero-yam-molmoact2.