[trainer, data] fix: do not drop_last SFT validation batches

[YeonwooSung]

What does this PR do?

Fixes #7464.

Both SFT trainers built the validation loader with train_batch_size and drop_last=True. A val set smaller than that batch (default 256 on Ray SFT) produced zero val batches. The trainer then did torch.mean(torch.tensor([])) and logged val/loss = NaN with exit code 0.

The config comment said micro_batch_size_per_gpu was also the val batch size, but that field was never used for val. PPO already uses drop_last=False and a dedicated val batch size.

This PR:

• Resolves val batch size as data.val_batch_size > data.micro_batch_size_per_gpu > len(val_dataset).
• Uses drop_last=False on the val sampler and dataloader.
• Skips logging val/loss when no val batches exist (SPMD path still all-reduces counts so ranks stay aligned).
• Adds data.val_batch_size to sft_trainer_engine.yaml.

Not a duplicate of #7401 (epoch-boundary resume) or #467 / #6319 (train-loss NaN).

Checklist Before Starting

• Search for similar PRs. Paste at least one query link here: https://github.com/verl-project/verl/pulls?q=is%3Apr+SFT+drop_last+validation
• Format the PR title as [{modules}] {type}: {description} (This will be checked by the CI)

Test

pytest tests/trainer/test_sft_val_batch_on_cpu.py -q

The helper was also executed via a direct module load on this machine (no torch/ray in the login Python). Ruff passed on the touched files.

API and Usage Example

# optional explicit val batch; otherwise micro_batch_size_per_gpu (default 4) is used
data.val_batch_size=8
data.val_files=/path/to/val.parquet

A 200-sample val set with train_batch_size=256 now runs validation instead of logging NaN.

Design & Code Changes

resolve_sft_val_batch_size() lives in verl/trainer/sft_val_utils.py so both the SPMD and Ray SFT trainers share one rule without importing each other.

Checklist Before Submitting

• Read the Contribute Guide.
• Apply pre-commit checks: pre-commit install && pre-commit run --all-files --show-diff-on-failure --color=always
• Add / Update the documentation.
• Add unit or end-to-end test(s) to the CI workflow to cover all the code. If not feasible, explain why: CPU unit tests cover the batch-size resolver.
• Once your PR is ready for CI, send a message in the ci-request channel in the verl Slack workspace.
• If your PR is related to the recipe submodule, please also update the reference to the submodule commit via git submodule update --remote or cd recipe && git pull origin main.

AI assistance was used to locate the bug and draft the patch. A human author reviewed every changed line.