ModelRetryMiddleware/ToolRetryMiddleware `Retry_on=SomeError` Retries Every Exception

The function `should_retry_exception` currently has `if callable(retry_on): return retry_on(exc)` before the `isinstance` branch. Exception classes are callable, so `retry_on(exc)` creates an exception instance, which is always truthy, making the function return `True` for all exceptions. The type-based branch is never reached for single exception classes.

Run the following Python snippet: `from langchain.agents.middleware._retry import should_retry_exception; print(should_retry_exception(KeyError('k'), ValueError))` -- this returns an exception instance (truthy) instead of `False`. Then using `ToolRetryMiddleware(max_retries=2, retry_on=TimeoutError)` with a tool that raises `KeyError` will retry the full budget instead of failing immediately.

This is a behavior change: code that previously relied on the buggy always-retry behavior will now fail fast as intended. Validate that all callers pass a type, tuple, or predicate; no other uses are expected. If custom callables depended on being invoked for every exception, they will no longer be called unless specified explicitly (which is correct). Rollback: revert to the original callable-first ordering if unforeseen issues arise, but this reintroduces the bug.