Fix train_batch_size and eval_batch_size to respect split_batches config

[MinuriRajapakse]

Fixes #45693

Problem

When split_batches=True is set in accelerator_config, the
train_batch_size and eval_batch_size properties were still
multiplying per_device_batch_size by n_gpu, which is incorrect.

When split_batches=True, the batch is split across devices rather
than replicated, so the total batch size equals per_device_batch_size
directly.

Fix

Added a check for split_batches in both train_batch_size and
eval_batch_size properties in TrainingArguments.

Testing

• Added a new test test_batch_size_respects_split_batches
• All 26 existing + new tests pass

[Rocketknight1]

cc @SunMarc