feat(sparserl-sync): integrate SparseRL-Sync sparse diff tracking into optimizer

What does this PR do?

Instrument the Megatron optimizer's in-place weight-copy sites so the
SparseRL-Sync package can compute per-param sparse-update indices for each
rollout. When sparse_update is not installed the imports fall back to
nullcontext no-ops and the upstream behaviour is fully preserved.

This PR is one of three coordinated changes:

• This PR — Megatron-LM: optimizer instrumentation
• slime-trainer: feat(sparserl-sync): integrate SparseRL-Sync sparse weight-sync
• scitix/sglang: feat(sparserl-sync): add sparse weight-sync path to ModelRunner

Modifications

megatron/core/optimizer/distrib_optimizer.py

• Best-effort from sparse_update import init_sparse_manager, sparse_diff_context; falls back to nullcontext / no-op when not installed.
• _build_model_and_main_param_groups: calls init_sparse_manager(model_param, shard_model_weight, shard_main_weight, param_range) for both fp16/bf16 and fp32 param groups so the SparseManager knows which DP-local buffer offset maps to which model param.
• _copy_main_params_to_model_params: wraps shard_model_param.data.copy_(shard_main_param) with sparse_diff_context(shard_model_param, shard_main_param) so the manager snapshots pre/post state and emits per-param sparse indices.

megatron/core/optimizer/cpu_offloading/hybrid_optimizer.py

• Best-effort from sparse_update import sparse_diff_context.
• Wraps both copy-back hooks (param_copy_back_gpu_hook H2D copy and fp32_param_copy_back_gpu_hook fp32→fp16 copy) with sparse_diff_context(target, source).

megatron/core/pipeline_parallel/p2p_communication.py

• Passes the group positional argument to all four torch.distributed.P2POp(...) constructors. It was previously omitted, which causes deadlocks when SparseRL-Sync's NCCL group overlaps with the world group during sparse weight-sync.

Pre-checks

• I want this PR in a versioned release and have added the appropriate Milestone
• I have added relevant unit tests
• I have added relevant functional tests
• I have added proper typing to my code
• I have added relevant documentation
• I have run the autoformatter.sh on my PR

Notes for reviewers

• All three import blocks are try/except ImportError; removing the
  sparse_update package from the environment restores byte-identical
  upstream behaviour with zero code changes.
• init_sparse_manager is a one-time bind per shard slice; it does not
  allocate any GPU memory itself.
• sparse_diff_context is a lightweight CUDA-stream-safe context manager; it
  does not add synchronisation points beyond the existing stream ordering.
• The p2p_communication.py fix is independent of SparseRL-Sync and is safe
  to cherry-pick separately.