Add nvFuser support for thunder.torch.randn

Nobody has claimed this yet.

## 🐛 Bug

```py
import thunder
import torch

@thunder.jit
def normal(shape, device):
return torch.randn(tuple(shape), device=device)

shape = [1, 4, 40, 84, 84]
t1 = normal(shape, device="cuda:0")
print(thunder.last_traces(normal)[-1]) # should have nvFusion0 there!
```
```py
def computation():
t6 = torch.randn((1, 4, 40, 84, 84), device=torch.device("cuda:0"), dtype=torch.float32) # t6: "cuda:0 f32[1, 4, 40, 84, 84]"
# t6 = ltorch.randn((1, 4, 40, 84, 84), generator=None, dtype=torch.float32, device=torch.device("cuda:0"), layout=torch.strided, requires_grad=False, pin_memory=False, out=None) # t6: "cuda:0 f32[1, 4, 40, 84, 84]"
# t6 = prims.randn((1, 4, 40, 84, 84), device=devices.Device("cuda:0"), dtype=dtypes.float32) # t6: "cuda:0 f32[1, 4, 40, 84, 84]"
return t6
```

cc @tfogal

Start by running the provided thunder.jit example with the stated shape and CUDA device, then inspect the resulting trace from thunder.last_traces(normal)[-1]. Follow how torch.randn is lowered through the shown ltorch and prims forms and compare the trace with the expected nvFusion0 output. Done means the trace includes nvFuser for this reproducible case.

Written by the indexing model from the issue text.