Add max_eval_batches argument to TrainingArguments

Description

Adds a max_eval_batches parameter to TrainingArguments that allows users to limit the number of batches used during evaluation.

Fixes #31561

Motivation

When working with large evaluation datasets, running evaluation on the entire dataset can be very slow. During development, hyperparameter tuning, or quick iteration, it's often sufficient to evaluate on a subset of the data.

This is similar to PyTorch Lightning's limit_val_batches parameter.

Changes

• ✅ Added max_eval_batches parameter to TrainingArguments
• ✅ Implemented batch limiting in Trainer.evaluation_loop
• ✅ Added test coverage
• ✅ Added documentation in parameter metadata

Usage Example

from transformers import Trainer, TrainingArguments

training_args = TrainingArguments(
    output_dir="./output",
    evaluation_strategy="steps",
    eval_steps=100,
    max_eval_batches=50,  # Only evaluate on 50 batches
)

trainer = Trainer(
    model=model,
    args=training_args,
    eval_dataset=eval_dataset,
)

# Evaluation will stop after 50 batches instead of going through entire dataset
trainer.evaluate()