no_grad is lost in jitted functions

Nobody has claimed this yet.

## 🐛 Bug

For a function decorated with `torch.no_grad`, the compile data of the jitted version has `is_grad_enabled` set to True when I would expect it to be False.

#### Code sample

```py
import torch
import thunder

@torch.no_grad
def f(x):
print(torch.is_grad_enabled())
return x * 2

jf = thunder.jit(f)

x = torch.ones((2,2))

f(x) # prints False
jf(x) # prints True
thunder.compile_data(jf).is_grad_enabled # True
```

Start by running the provided Python example and compare f(x) with thunder.jit(f)(x), including thunder.compile_data(jf).is_grad_enabled. Then inspect the thunder.jit and compile_data entry points for how torch.no_grad state is carried into compiled functions; done means the jitted call and compile data report gradients disabled.

Written by the indexing model from the issue text.