Fix action_loss_per_sample shape for has_real_action

[Loki-Lu]

This PR is used to fix: #25

[jeqcho]

Independent confirmation: we hit the identical crash fine-tuning DreamZero-AgiBot on bimanual YAM data with per_device_train_batch_size=4 (the yam_training.sh default), and arrived at this exact one-line fix before finding this PR. has_real_action[:, None] right-aligns as (1, B, 1) against the [B, T, D] loss; B=1 happens to broadcast with correct semantics, which is why batch-1 runs never see it — but it also means the released batch-4 YAM script cannot have run as shipped.

We ran with this fix for a full 12k-step LoRA fine-tune at global batch 32 (training stable throughout; released at https://huggingface.co/robocurve/dreamzero-yam-molmoact2, training code at https://github.com/robocurve/dreamzero-yam). Would be great to see this merged — it currently blocks anyone from using batch > 1 on the LeRobot fine-tuning path.