fix(dist): preserve new_group options across reloadable group reload

[EazyReal]

What changed

ReloadableProcessGroup now captures the original positional args and kwargs passed to torch.distributed.new_group and replays them verbatim in reload_process_groups(), instead of recreating each group with only old_new_group(ranks=..., backend="nccl").

The gloo-skip and rank-extraction logic in the new_group monkeypatch is factored into _is_gloo_group() and _get_group_ranks() helpers (behavior-preserving; also removes the operator-precedence ambiguity in the old a and b or c and d gloo check).

Adds a CPU regression test (test_reload_process_groups_preserves_new_group_timeout) that drives a create -> destroy -> reload cycle and asserts timeout and pg_options survive the reload, and registers the test file in the cpu-unittest CI list.

Why

On offload/reload, reload_process_groups() recreated each NCCL group with only ranks and backend="nccl", dropping the original timeout, pg_options, and any other construction options. A job that launches with a longer distributed timeout would create its initial NCCL groups correctly but silently fall back to PyTorch's default timeout after a reload, so long-running actor train/update collectives could trip the NCCL watchdog even though the operator configured a larger timeout. The reload lifecycle is owned by this wrapper, so the wrapper is the only place that can preserve and replay the original new_group contract.

Gloo and single-rank groups keep their existing bypass (returned directly, not wrapped/registered).

Validation

CPU-only, no GPU required:

• pytest tests/test_reloadable_process_group_memory_check.py -q -- the new test fakes dist.new_group/old_new_group, runs new_group -> destroy_process_groups -> reload_process_groups, and asserts the reloaded inner group is recreated with the same ranks, backend, timeout, and pg_options objects. It fails against the pre-fix code (which hard-codes ranks=..., backend="nccl").
• The test file is registered in .github/workflows/pr-test.yml.j2's cpu-unittest job so it runs in CI.

[EazyReal]

@zhuzilin could you review this one? Reloadable distributed groups were recreating with default new_group options after reload; this preserves the original group options so timeout/backend semantics survive reload.