add action mask in loss computation

[HomerIsAFool]

Motivation

In the implementation of π pytorch version, the loss is computed in the forward() as below:
https://github.com/Physical-Intelligence/openpi/blob/main/src/openpi/models_pytorch/pi0_pytorch.py#L373

return F.mse_loss(u_t, v_t, reduction="none")

But the dim of u_t and v_t are max_dim from the config (32 in π05). This means that the padded actions will also be included in the loss calculation. I think this may harm the model performance.

The lerobot π05 has fixed this by using action_mask when calculating the loss:
https://github.com/huggingface/lerobot/blob/main/src/lerobot/policies/pi05/modeling_pi05.py#L1257

Modification

In this PR, we create a new param actual_action_dim in src/openpi/models/pi0_config.py to represent the actual action dim.
And create a new param action_mask in class PI0Pytorch.

In the forward func, use the action_mask to eliminate the impact of padded actions on the loss.

return F.mse_loss(u_t, v_t, reduction="none") * self.action_mask