feat(dspark): add offline USP training support

[calyoung80]

Motivation

DSpark offline training currently processes the full sequence on every draft
sequence-parallel rank, which limits the usable sequence length and duplicates
target hidden-state storage and compute.

This PR adds a USP-backed offline training path for DSpark. It shards sequence
tensors across the draft-SP group while preserving global sequence semantics
for draft attention and loss computation.

Modifications

• Add rank-local DSpark offline sample normalization for USP:
  • shard input_ids, loss_mask, hidden_states, and
    target_last_hidden_states;
  • pad the final shard to a static sequence length for collective
    communication;
  • carry a local attention_mask through the collator and training strategy.
• Add an autograd-correct sequence all-gather helper:
  • reconstruct global projected target hidden states for draft attention;
  • route gradients back to the owner rank with reduce_scatter_tensor;
  • retain a Gloo-compatible reference fallback.
• Add usp as a DSpark attention backend:
  • use SDPA only for local Transformers draft-model construction;
  • keep USP topology and sequence communication under SpecForge control;
  • avoid importing CUDA-only sequence-parallel dependencies for this path.
• Project target hidden states before the HCCL exchange to reduce communication
  width.
• Synchronize valid anchor width across draft-SP ranks and support empty local
  shards safely.
• Add focused coverage for data sharding, padding, collator propagation,
  distributed gather/backward behavior, backend mapping, and empty-shard anchor
  handling.

Related Issues

N/A

Accuracy Test

This PR changes DSpark training infrastructure rather than model architecture
or inference kernels.

Accuracy validation has not been run yet. A multi-rank DSpark offline training
comparison against the non-USP baseline is required before merge.

Benchmark & Profiling

This PR is intended to enable longer offline training sequences and reduce
per-rank sequence-state memory.

Benchmark and profiling results have not been collected yet. Suggested
follow-up measurements:

• maximum trainable sequence length;
• peak device memory per rank;
• HCCL communication time;
• tokens/s versus the non-USP DSpark offline baseline.

Checklist

• Format code with pre-commit.
• Add focused unit tests.
• Run the focused test suite:
  pytest tests/test_algorithms/test_dspark_usp.py tests/test_algorithms/test_model_providers.py tests/test_utils/test_dflash_losses.py
• Update documentation / examples as needed.
• Provide multi-rank accuracy and throughput / latency results before merge.
• Add a DCO sign-off:
  Signed-off-by: calyoung80 <zsucal@hotmail.com>

[gemini-code-assist]

Caution

The consumer version of Gemini Code Assist on GitHub has been sunset. All code review activity has officially ceased.

[jiapingW]

Hi, can you add your results of dspark draft model that trained on long sequence?