Support tracing through input-dtype asserts

Nobody has claimed this yet.

## 🚀 Feature

When there are asserts on the input dtypes, we would want to pass the actual dtype instead of jitting with the thunder dtype equivalent

```python
import torch
import thunder

a = torch.rand((2, 2), dtype=torch.bfloat16)

def foo(x):
assert x.dtype in (torch.float32, torch.bfloat16), x.dtype
return x

jfoo = thunder.jit(foo)

b = jfoo(a)
```

results in the following error:

```shell
File "lightning-thunder/repro.py", line 7, in foo
assert x.dtype in (torch.float32, torch.bfloat16), x.dtype
^^^^^^^^^^^^^^^^^^
AssertionError: thunder.dtypes.bfloat16
```

Start by running the reproduction in lightning-thunder/repro.py and trace how input dtypes are represented during thunder.jit tracing. Look for the dtype handling involved in evaluating the assert, then verify that the example accepts torch.bfloat16 while preserving the asserted dtype behavior.

Written by the indexing model from the issue text.