how to provide overrides needed for model compatibility

@tfogal is already working on this.

Since Jul 22, 2024.

While transformers BERT is not yet fully working, we are getting closer.
However, there is the need to disable some data-dependent control flow (typically: checks) to get it to work.
Transformers itself hides some but not all behind a check for compiling, e.g.

https://github.com/huggingface/transformers/blob/0fdea8607d7e01eb0e38a1ebeb7feee30a22f0cf/src/transformers/modeling_attn_mask_utils.py#L256-L260
```
is_tracing = (
torch.jit.is_tracing()
or isinstance(inputs_embeds, torch.fx.Proxy)
or (hasattr(torch, "_dynamo") and torch._dynamo.is_compiling())
)
```

So here is a candidate for how we might currently be able to run BERT after fixing #805 (and subsequent bugs):
```
import transformers, thunder, torch
@thunder.core.jit_ext.register_general_jit_lookaside(
transformers.modeling_utils.PreTrainedModel.warn_if_padding_and_no_attention_mask
)
@thunder.core.jit_ext.interpreter_needs_wrap
def dummy(*args):
pass

@thunder.core.jit_ext.register_general_jit_lookaside(
torch._dynamo.is_compiling
)
@thunder.core.jit_ext.interpreter_needs_wrap
def is_compiling():
return True

m = transformers.BertForSequenceClassification(transformers.BertConfig())
inp = torch.randint(1, 20, (1, 32))
jm = thunder.jit(m)
jm(inp)
```

We might submit an issue to transformers to check for is_tracing in `transformers.modeling_utils.PreTrainedModel.warn_if_padding_and_no_attention_mask`
but in general I wonder how to provide the such compatibility lookasides to users.

possible variants:
- A compat executor,
- detecting the module we are tracing and adding them just for transformers (too much magic),
- always having the `torch._dynamo.is_compiling`-lookaside?
...