Disable gradient checkpointing.

@func: load the state of the optimizer and scaler and resume the training progress

Constructs pre-processor and post-processor pipelines for the PI0 policy. The pre-processing pipeline prepares input data for the model by:

Normalize to [-1, 1]. If using quantiles, clips outliers to [q01, q99] before scaling.

Select a single action given environment observations.

@func: return a state dictionary for the optimizer and scaler

Warning: does not synchronize the deque!

This step does not alter the feature definitions.

Unnormalize from [-1, 1] back to original scale. No clipping during unnormalization (allow reconstruction to exceed bounds).