remove old-style fsdp/ddp

[t-vi]

cc: @crcrpar

Plan:

Replace with transform-aware module functions:

• thunder/tests/distributed/test_checkpoint.py::DistributedCheckpointTest::test_get_model_state_dict
• thunder/tests/distributed/test_checkpoint.py::DistributedCheckpointTest::test_load_model_state_dict

Removed as not (currently) applicable

• FAILED thunder/tests/distributed/test_ddp.py::DDPTest::test_ddp_model_as_argument
  --> We are not re-entrant.

Fixed

• thunder/tests/distributed/test_ddp.py::DDPTest::test_ddp_grad_parity_with_without_bucketing * 2
• thunder/tests/distributed/test_ddp.py::test_native_ddp * 4
• test_fsdp.py::FSDPTest::test_fsdp_with_no_sync_grad_accumulation * 4 - Thank you @crcrpar

[crcrpar]

I'm speculating that the failure of fsdp no_sync parity tests is caused by the mismatch between the parameters that _sync_grads collects and ones that _stash_grad_for_fsdp_prim_impl attaches the unsharded grads.

• lightning-thunder/thunder/distributed/__init__.py

  Lines 157 to 159 in 5e18c2e

  	params_with_grad = tuple(filter(lambda p: hasattr(p, "_thunder_fsdp_unsharded_grad"), module.parameters()))
  	if not params_with_grad:
  	return

  .

• lightning-thunder/thunder/executors/torchex.py

  Lines 2116 to 2128 in 5e18c2e

  	def _stash_grad_for_fsdp_prim_impl(
  	grad: torch.Tensor,
  	param_fqn: str,
  	compile_data: CompileData,
  	) -> None:
  	grad_name = "_thunder_fsdp_unsharded_grad"
  	param = compile_data.fn.get_parameter(param_fqn)
  	if torch.is_tensor(unsharded_grad := getattr(param, grad_name, None)):
  	unsharded_grad += grad
  	else:
  	setattr(param, grad_name, grad)
  	return grad

[lantiga]

Stamped!