fix: sync DeepSpeed gradient_accumulation_steps from TrainPipelineConfig

[shuheng-liu]

What this does

Makes TrainPipelineConfig.gradient_accumulation_steps the single source of truth for gradient accumulation. Previously, if the value in TrainPipelineConfig differed from the one in the Accelerate/DeepSpeed YAML, training would raise ValueError — forcing duplicate edits in two places. Now, when DeepSpeed is the distributed backend, the DeepSpeed config's gradient_accumulation_steps is overridden at runtime to match TrainPipelineConfig, with a main-process-only logging.warning when it actually changes a value.

Key details:

• New helper _sync_deepspeed_gradient_accumulation_steps(accelerator, cfg) in src/opentau/scripts/train.py runs on all ranks (_prepare_deepspeed reads the config on every rank) and before encode_accelerator_state_dict + init_trackers, so the wandb-logged accelerator config also reflects the overridden value.
• gradient_accumulation_steps is now always passed to Accelerator(...) (even when ==1) for explicit intent.
• Non-DeepSpeed backends are a no-op.

Label: (🐛 Bug)

How it was tested

Added tests/scripts/test_train.py with 4 unit tests for the helper:

1. Non-DeepSpeed backend → no-op, no warning.
2. DeepSpeed, YAML matches cfg → no-op, no warning.
3. DeepSpeed, mismatch on main process → overrides both hf_ds_config.config and deepspeed_plugin.gradient_accumulation_steps, emits one WARNING containing both values.
4. DeepSpeed, mismatch on non-main rank → override still happens, no warning (log-once behavior).

All four pass locally. Full pre-commit suite (ruff, ruff-format, bandit, license header, typos, pyupgrade, secrets) passes on the changed files.

How to checkout & try? (for the reviewer)

git fetch origin claude/sync-grad-accum-from-train-config
git checkout claude/sync-grad-accum-from-train-config
pytest -sx tests/scripts/test_train.py

To exercise the bug scenario end-to-end (requires a DeepSpeed-enabled environment):

# configs/examples/accelerate_deepspeed_config.yaml has gradient_accumulation_steps: 2
accelerate launch --config_file configs/examples/accelerate_deepspeed_config.yaml \\
  src/opentau/scripts/train.py --config_path=<your_config.json> \\
  --gradient_accumulation_steps 4
# Previously: ValueError. Now: starts cleanly with a warning about the override,
# and DeepSpeed's startup banner reports gradient_accumulation_steps=4.

Checklist

• I have added Google-style docstrings to important functions and ensured function parameters are typed.
• My PR includes policy-related changes.
  • If the above is checked: I have run the GPU pytests (pytest -m "gpu") and regression tests.