DTensor: Get it working on the thunderfx path when requires_grad=True.

@kshitij12345 is already working on this.

Since Jun 13, 2025.

This currently doesn't work.

```python
from torch.distributed.tensor import DTensor
from torch.distributed import init_device_mesh
from thunder.dynamo import thunderfx
import torch
import os

# Create mesh
os.environ["RANK"] = "0"
os.environ["LOCAL_RANK"] = "0"
os.environ["WORLD_SIZE"] = "1"
os.environ["MASTER_ADDR"] = "localhost"
os.environ["MASTER_PORT"] = "29500"
mesh = init_device_mesh("cuda", (1,), mesh_dim_names=["i"])

hidden_size = 16

def model(x, w):
return torch.mul(x, w)

weight = DTensor.from_local(torch.randn(hidden_size, hidden_size, requires_grad=True), device_mesh=mesh)
in_dtensor = DTensor.from_local(torch.randn(hidden_size, hidden_size, requires_grad=True), device_mesh=mesh)

expected = torch.compile(model)(in_dtensor, weight)

tmodel = thunderfx(model)
actual = tmodel(in_dtensor, weight)

torch.testing.assert_close(actual.to_local(), expected.to_local())
try:
assert tmodel._backend.subgraph_infos[0].split_reasons == [] # Fails
finally:
torch.distributed.destroy_process_group()
```
Split reasons:
```py
In [2]: tmodel._backend.subgraph_infos[0].split_reasons
Out[2]: [SplitReason(reason_type=, info='Failed while running meta for node with name: mul and target: , see exception field', exception='aten.mul.Tensor: got mixed torch.Tensor and DTensor, need to convert all torch.Tensor to DTensor before calling distributed operators!')]
```