truncated_bptt_steps should be a property of the LightningModule, not the Trainer

[ananthsub]

🚀 Feature

Very similar to #4317

https://pytorch-lightning.readthedocs.io/en/latest/common/trainer.html#truncated-bptt-steps
It's highly architecture dependent, meaning it's closer to the research code than the engineering code. Accordingly, this should live on the LightningModule, not on the Trainer.

Pitch

Define a property method for truncated_bptt_steps inside the LightningModule

Changes to the LightningModule

self._truncated_bptt_steps: Optional[int] = None

@property
def truncated_bptt_steps(self) -> Optional[int]:
    return _truncated_bptt_steps

@truncated_bptt_steps.setter
def truncated_bptt_steps(self, tbptt_steps) -> None:
    self._truncated_bptt_steps = tbptt_steps

Then the training loop can read these settings off the lightning module and adjust accordingly, similar to what's done today for automatic/manual optimization. We can then deprecate the trainer flag for this setting.