Fix metric padding removal after skip_first_batches

[Excelius-Wang]

What does this PR do?

skip_first_batches wraps the batch sampler in SkipBatchSampler, which does not expose batch_size. This makes DataLoaderShard.total_batch_size raise; the suppressed exception leaves remainder=-1, so gather_for_metrics keeps duplicated padding samples in the final batch.

With 10 samples, two processes and a per-process batch size of 2, normal iteration gathers 10 samples. Calling skip_first_batches(loader, 0) instead gathers 12, including duplicated samples 0 and 1.

Unwrap skip samplers before reading the existing batch-size metadata, and preserve _drop_last when reconstructing the loader. This also handles nested skipping without changing sampling order.

Validation

• 66 related tests passed, 16 skipped.
• All 18 new parameterized regressions fail on the baseline and pass with the fix.
• Actual CPU/Gloo testing: the expanded 48-case two-process matrix passed twice; the earlier 24-case three-process matrix also passed twice. Coverage includes split batches, different batch sizes, divisible and non-divisible dataset sizes, drop_last, and zero/one/nested skips.
• The distributed regression fails on the baseline.
• Full make quality and git diff --check pass.

GPU/TPU execution was not tested. Changes were prepared with AI assistance and reviewed by the contributor.

Before submitting

• Read the contribution guidelines.
• Added regression tests.
• Prior issue discussion: none.
• Documentation changes: not needed; this restores existing behavior.