fix: use sys.modules.get() to avoid KeyError in modeling_utils

[Krishnachaitanyakc]

Summary

Fixes #45003

_can_set_attn_implementation and _can_set_experts_implementation in PreTrainedModel use sys.modules[cls.__module__], which raises KeyError when a module has been removed from sys.modules at runtime. This can happen with:

• vLLM CI (modules deleted during cleanup)
• Griptape Nodes and similar tools that manipulate sys.modules for dependency management
• Any lazy-loading or packaging tool that removes modules from the registry

Changes

• Replace sys.modules[cls.__module__] with sys.modules.get(cls.__module__) (2 locations)
• Add class_module is None guard before the existing hasattr(class_module, "__file__") check
• When the module is missing, returns False — the same safe fallback already used for modules without __file__ (e.g., Jupyter notebooks)

Why not a broader fix?

This is intentionally minimal (4 lines changed). The existing comment already acknowledges that module lookup can fail ("This can happen for a custom model in a jupyter notebook or repl"), and sys.modules.get() is the standard safe-access pattern.

Test plan

• Verified the fix matches the approach discussed by maintainers in modeling_utils unsafely accesses sys.modules[] #45003
• Confirmed no open PRs exist for this issue
• sys.modules.get() returns None (not KeyError) when module is absent — None check short-circuits to return False
• Normal-case behavior is unchanged (module exists → same flow as before)

AI assistance was used in preparing this PR.