refactor: share training batch production

[tim0120]

What

Extracts the trainer-bound batch lifecycle behind a source-independent interface:

• TrainingBatchSource prepares examples and handles source-specific post-send work.
• TrainingBatchProducer is the single loop that publishes TrainingBatch, advances the trainer-aligned step, and notifies the source.
• The online rollout orchestrator and static dataset source both use that loop.

This is a stacked architectural draft on #2954 for examining whether this is the right invariant boundary.

Why

#2954's DatasetBatchProducer still owned its own send/step loop, making it structurally close to a renamed static orchestrator. The first interface shared by rollout training and static SFT is the prepared trainer batch, not environments, dataset rows, or rollouts.

This change shares the lifecycle at that boundary while leaving source-specific behavior on the appropriate side:

• Rollout draining, eval handling, traces, and rollout metrics remain in Orchestrator.
• Dataset rendering, cursor accounting, trainer acknowledgement, and static metrics remain in DatasetBatchSource.

No config, transport, trainer, normalization, or checkpoint format changes are intended.

Validation

• Ruff check and format check passed for the three changed files.
• Python compilation passed for the three changed files.
• A focused in-memory producer check verified publication order, step advancement, and the post-send callback contract.
• git diff --check passed.

No README changes or new test files.