Fix TorchDynamo crash in StaticCache by validating offloading and offload_only_non_sliding arguments

[Flakes342]

What does this PR do?

Fixes #41164

Before submitting

• This PR fixes a typo or improves the docs (you can dismiss the other checks if that's the case).
• Did you read the contributor guideline,
  Pull Request section?
• Was this discussed/approved via a Github issue or the forum? Please add a link
  to it if that's the case.
• Did you make sure to update the documentation with your changes? Here are the
  documentation guidelines, and
  here are tips on formatting docstrings.
• Did you write any new necessary tests?

Who can review?

@ArthurZucker @Rocketknight1

Description

This PR addresses a bug in StaticCache that can lead to torch.compile / TorchDynamo crashes when users pass:

• Unexpected keyword arguments (**kwargs)
• Misordered positional arguments
• Non-boolean values for offloading or offload_only_non_sliding

Root causes / API issues:

1. Positional flexibility vs. keyword expectations

   • StaticCache accepts config, max_cache_len as positional arguments, but offloading and offload_only_non_sliding were also sometimes passed positionally.
   • This can silently assign non-boolean types (e.g., a device object) to offloading, leading to CUDA/TorchDynamo crashes in update().

2. Inconsistent argument handling across cache classes

   • Base Cache is strict: offloading and offload_only_non_sliding must be booleans.
   • StaticCache had a flexible signature with **kwargs, which could allow invalid arguments without immediate error.
   • Deprecated children sometimes passed extra arguments through **kwargs, increasing risk of silent misassignment.

3. Potential user misuse

   • Users passing device/dtype arguments positionally (common in examples or from auto-generated code) could crash torch.compile unexpectedly.
   • Without clear error messages, this was non-obvious and difficult to debug.

Reproduction

Using Qwen3-omni or similar models:

• Here, device and compute_dtype were passed as positional arguments.
• offloading is expected to be a bool, but now receives "cuda:0".
• This triggers the crash: InternalTorchDynamoError: AttributeError: 'int' object has no attribute 'device'

Solution

1. Validate types in StaticCache.__init__:
   Ensure offloading and offload_only_non_sliding are bool.
   Raise a clear TypeError if invalid.

2. Catch unknown kwargs:
   Raise TypeError if users pass unrecognized keyword arguments (**kwargs) to prevent silent misassignment.

3. Forward **kwargs safely in children:
   Deprecated children like OffloadedHybridCache are untouched for now, but any kwargs they pass are safely handled in StaticCache.

Additional Notes

• We could optionally enforce keyword-only arguments via * to fully prevent positional misassignment, but this may break backward compatibility.
• Deprecated cache classes (OffloadedHybridCache, etc.) are left untouched since they will be removed in v4.59.
• A more advanced fix could include runtime coercion of devices to offloading=False to fully prevent user mistakes, but raising errors is safer and explicit.

[ArthurZucker]

sorry I don't see how someone would want to just pass 0 like that?