[cache] @cache raises TypeError for unhashable arguments (e.g. Pydantic model inputs) #59

Summary

In nitrostack/core/additional_decorators.py, @cache builds its cache key directly from positional and keyword arguments. When an argument is unhashable, such as a non-frozen Pydantic v2 model, the cache lookup raises TypeError instead of executing the function.

Findings

• cache() filters out ExecutionContext arguments but retains other arguments in their original form.
• The cache key is constructed using tuple(filtered_args) and frozenset(filtered_kwargs.items()). Both structures require their contents to be hashable.
• When a Pydantic model is passed as an argument, checking whether the key exists in cache_store can raise TypeError: unhashable type.
• This affects cached tool methods that receive Pydantic input models and other functions that accept unhashable arguments.
• The current implementation does not provide a custom key function or a clear error explaining which argument caused the failure.

Proposal

1. Serialize Pydantic BaseModel arguments using model_dump(mode="json") and generate a deterministic cache key from the serialized data.
2. Add an optional key= callable to cache() so developers can define cache keys for custom argument types.
3. Raise a clear, descriptive error identifying the unhashable argument and its type when a valid cache key cannot be constructed.
4. Add tests covering Pydantic models, unhashable positional and keyword arguments, and custom key functions.