Excessive caching for `torch.dtype` objects

🐛 Bug

It seems that when passing a torch.dtype as argument of a jitted function, Thunder caches the value and the resulting rerun of the function outputs wrong values. For example:

To Reproduce

The following code:

import torch
import thunder

def to_dtype(a, dtype):
    return a.to(dtype)

jit_dtype = thunder.jit(to_dtype)

a = torch.randn((1, 2), dtype=torch.float32)

print(jit_dtype(a, torch.bfloat16))
print(jit_dtype(a, torch.float16))
print(jit_dtype(a, torch.float64))

Will print 3 times a tensor of dtype bf16.

Code sample

Expected behavior

I expect the dtype to change every time, not just once.

Additional context

The issue can be solved by temporarily disabling cacheing (i.e. cache=NO_CACHING argument to thunder.jit).