[WIP] Fix realtime entropy patching

[Vectorrent]

This is a followup to the comment made here.

I want to load and train an entropy model alongside of the latent model - rather than loading it from checkpoint. The current code does not support this, so I added an additional PatcherArgs attribute, called entropy_model. Here, you can pass a Pytorch module directly, instead of a checkpoint path.

I also removed a self.output_proj variable, because it crashes when using entropy patching. The entropy_model_checkpoint_dir attribute does not exist on LocalModelArgs. Regardless, this variable is unused - and probably not necessary?

Finally, I removed a logger warning that was extremely spammy. We are already gating this warning with the BLT_SUPPRESS_ATTN_ERROR environment variable, and I'm assuming you didn't intend on throwing the same warning with every forward pass. I'm sure this was overlooked, and you meant to remove this extra logging.

Let me know if you have any questions!

[Vectorrent]

I noticed that the Patcher expects a static batch size to be set, but my custom model is more dynamic than that - and often uses a variable batch size. Thus, I need to adjust the batch size in the patcher in realtime, during training. I made that possible with an override in the forward method.

[Vectorrent]

I reverted some of the Patcher code to it's original style, since I realized I could just use the entropies argument to handle the batch size stuff.

I also fixed a warning here, since passing entropies in that way was causing this:

bytelatent/data/patcher.py:535: UserWarning: To copy construct from a tensor, it is recommended to use sourceTensor.clone().detach() or sourceTensor.clone().detach().requires_grad_(True), rather than torch.tensor(sourceTensor).
  scores = torch.tensor(entropies, dtype=torch.float32)