Support FSDP and torch.compile

Nobody has claimed this yet.

## 🚀 Feature

### Motivation

We use `async_op=True` to reorder the collective calls in the trace, but that's not supported by Dynamo:

```python
[rank7]: torch._dynamo.exc.Unsupported: CollectiveFunctionRewriteVariable can't support async_op=True for

[rank7]: from user code:
[rank7]: File "thunder.torch_interpreted_func_94", line 13, in torch_interpreted_func
[rank7]: p8 = torch_all_gather_prim_impl(t2, _torch_distributed_distributed_c10d_ProcessGroup_0, True) # p8: "FUTURE cuda:7 bf16[2560, 2048]"
[rank7]: File "/home/carlos/lightning-thunder/thunder/executors/torchex.py", line 1687, in _all_gather_prim_impl
[rank7]: handle: None | torch.distributed.distributed_c10d.Work = torch.distributed.all_gather_into_tensor(
```

### Pitch

We could support this two ways:

a. Disable `fullgraph=True` to allow graph-breaks via #281
b. Do not let the torch_compile_ex to fuse through these collective calls.

Option b) is what I would recommend.

### Alternatives

Do not support this.

### Additional context

Requires #140 to land first.

cc @carmocca @awaelchli @crcrpar @apaz-cli

Start with thunder/executors/torchex.py around the _all_gather_prim_impl entry point and review prerequisite issue #140. Compare the proposed graph-break and non-fusion approaches for async collective calls, then identify the FSDP and torch.compile scenarios that would demonstrate completion.

Written by the indexing model from the issue text.