…NVIDIA-NeMo#2229)

Several Megatron-LM infrastructure knobs were hard-coded inside
``nemo_rl/models/megatron/setup.py`` even though the underlying
framework already supports them as configurable. Users running on
varied storage backends (NFS) or training very large models (e.g.
Nemotron 3 Super 120B) had to patch the source file to override them.

Expose the following four knobs as ``NotRequired`` fields on
``MegatronConfig``:

* ``distributed_timeout_minutes`` -- forwarded as ``timeout`` to
  ``torch.distributed.init_process_group``. The default NCCL timeout
  (~10 minutes) is too short for initial weight conversion of very
  large models on slow storage.
* ``async_save`` -- enables non-blocking checkpoint writes via
  Megatron-Core's existing async save plumbing.
* ``check_for_nan_in_grad`` -- toggles Megatron-LM's per-step
  NaN/Inf grad check; useful for debugging but adds non-trivial
  overhead in steady-state training.
* ``logging_level`` -- forwarded to Megatron-LM's ``LoggerConfig``.

When any key is absent the historical hard-coded behaviour is
preserved, so existing user configs continue to work unchanged.

Plumbing:

* ``setup_distributed()`` now optionally accepts the policy config
  and threads ``distributed_timeout_minutes`` through to
  ``init_process_group``.
* ``_create_checkpoint_config()`` accepts the policy config and reads
  ``async_save``.
* ``_create_megatron_config()`` reads ``check_for_nan_in_grad`` and
  ``logging_level`` from the existing ``config`` it already receives.
* ``MegatronPolicyWorker`` passes ``config`` to ``setup_distributed``.

Coverage:

* New ``TestSetupDistributedTimeout`` exercises the three timeout
  branches (no config / config without key / config with key).
* ``TestCreateCheckpointConfig`` gains three tests covering the
  ``async_save`` default and override behaviour.

Out of scope (intentionally left for a follow-up to keep this PR
reviewable): ``fully_parallel_save``, ``fully_parallel_load``, and
``load_rng``. These are referenced in two places (the main and
reference checkpoint configs) and changing them could affect
convergence-test behaviour, so they deserve their own focused PR.

Refs: NVIDIA-NeMo#2229
Signed-off-by: lonexreb <[email protected]>

…#2229)

Address review feedback on PR NVIDIA-NeMo#2391: the previous

    megatron_cfg = config["megatron_cfg"] if config and "megatron_cfg" in config else {}
    async_save = megatron_cfg.get("async_save", False) if megatron_cfg else False

was redundant — when ``megatron_cfg`` is the empty dict, ``.get()``
already returns the default, and the ``if megatron_cfg else default``
chain only differs from a bare ``.get()`` when ``megatron_cfg`` is
falsy, in which case it is the empty dict and ``.get()`` returns the
same default. Simplify to four bare ``.get()`` calls and switch the
``config`` truthiness check to an explicit ``config is not None``,
matching the pattern in PR NVIDIA-NeMo#2390 and aligning the two PRs' style.

Behaviour is identical: defaults match the historical hard-coded
values when keys are absent, and explicit user values flow through.
The existing tests in ``TestCreateCheckpointConfig`` continue to
exercise both branches.

Refs: NVIDIA-NeMo#2229
Signed-off-by: lonexreb <[email protected]>