fix(rollout): make dynamic refills granular

[EazyReal]

Problem

Dynamic sampling currently requires over_sampling_batch_size to be at least rollout_batch_size, even though the rollout state machine supports any positive refill granularity. A single rejected prompt group can therefore launch a full replacement wave.

At rollout completion, non-partial generation tasks are also left running while SGLang is drained. Multi-turn custom generation can issue new requests during that drain, delaying or preventing it from reaching idle.

Change

• accept any positive over_sampling_batch_size
• cancel and await non-partial generation tasks before draining SGLang
• preserve partial-rollout prefixes through server drain
• make the complete abort lifecycle exception-safe while re-raising the original failure
• align English and Chinese dynamic-sampling documentation
• add CPU regressions for validation, cancellation ordering, partial preservation, and failure cleanup

The existing scheduler remains unchanged: rejected groups decrement accepted-plus-pending coverage, and the configured refill granularity restores it.

Validation

• python tests/test_sglang_rollout_abort.py -q — 4 passed
• focused Python 3.10 suite — 33 passed
• python .github/workflows/generate_github_workflows.py
• pre-commit run --all-files
• git diff --check

No live GPU or distributed rollout was run; the affected lifecycle boundaries are covered with CPU and mocked regressions.