Fix evaluation loop behavior when max_eval_steps is set

[PRATHAMESHG1234]

This PR follows up on #41299 and fixes remaining issues around max_eval_steps.

• Adds max_eval_steps to TrainingArguments
• Limits evaluation by number of batches
• Fixes num_samples handling for iterable datasets
• Preserves callback event ordering

All trainer tests pass locally.

[github-actions]

View the CircleCI Test Summary for this PR:

https://huggingface.co/spaces/transformers-community/circle-ci-viz?pr=42980&sha=ed59aa

[PRATHAMESHG1234]

The tests_torch failure is due to a known ResNet initialization bug (see #31841), not this PR.

Failing test: tests/models/resnet/test_modeling_resnet.py::ResNetModelTest::test_can_load_ignoring_mismatched_shapes
Error: AssertionError: 0.11898990720510483 not less than or equal to 0.1 for classifier.1.bias

This is a pre-existing issue where ResNet's ignore_mismatched_sizes=True produces out-of-bound initialization values. My changes only modify evaluation loop behavior (max_eval_steps) and don't touch model initialization code.

All other tests passed. Requesting maintainer review.

[PRATHAMESHG1234]

@sgugger @ArthurZucker @muellerzr Could you review this PR? The ResNet test failure appears unrelated to the changes.