Add DeepSeek V3 support

[SamJSui]

Summary

Adds Liger Kernel support for Hugging Face Transformers deepseek_v3 / DeepSeek V3. This is distinct from the existing deepseek_v32 support work.

• Adds a DeepSeek V3 fused linear cross-entropy forward path.
• Wires apply_liger_kernel_to_deepseek_v3 into public exports and model-type autodispatch.
• Patches RMSNorm, dense and shared-expert SwiGLU, CrossEntropyLoss, and FusedLinearCrossEntropy for both class-level and instance-level patching.
• Adds bf16/fp32 convergence coverage for fused-loss and materialized-logits paths.
• Adds direct API regressions for tuple returns, logits_to_keep, and explicit shift_labels.
• Updates the README patching table.

Closes #623.

Details

DeepSeek V3 uses partial, interleaved RoPE that is incompatible with Liger's generic RoPE swap. Passing rope=True therefore emits a warning and leaves both upstream DeepSeek V3 RoPE functions unchanged.

The routed expert implementation also remains upstream. Liger patches dense MLPs and sparse-layer shared_experts with the existing Qwen3 MoE-compatible SwiGLU wrapper; a local Triton 3.2 bf16 fused-MoE backward probe failed because atomic_add does not support bf16.

The FLCE forward follows the current Transformers API contract, including @can_return_tuple, logits_to_keep, explicit shift_labels, and the shared PEFT/FSDP-safe LM-head loss helper.

Testing Done

• Hardware Type: NVIDIA GeForce RTX 4070 Ti SUPER
• run make test to ensure correctness
• run make checkstyle to ensure code style
• run make test-convergence to ensure convergence

The checked boxes mean the commands were run. The two broad GPU targets did not exit zero locally; every nonzero case reproduced unchanged on clean origin/main at 780e76b, while all DeepSeek V3 checks passed.

Clean validation environment: Python 3.10.12, PyTorch 2.13.0+cu130, Triton 3.7.1, Transformers 5.15.0.

make test
# 3998 passed, 1149 skipped, 14 xfailed
# 2 bf16 numeric-tolerance failures:
# - test_geglu.py (reproduces on origin/main)
# - test_poly_norm.py (reproduces on origin/main)

make test-convergence
# 158 passed, 9 skipped, 3 xfailed, 1 xpassed
# 5 non-DeepSeek failures, all reproduced on origin/main:
# - fp32 mini_llava (with and without materialized logits)
# - bf16 mini_llama4
# - bf16 mini_qwen3_5_moe multimodal
# - bf16 mini_qwen3_moe with materialized logits
# All 4 DeepSeek V3 convergence cases passed.

make checkstyle
# passed

git diff --check
# passed

Focused checks completed successfully:

PYTHONPATH=src python3 -m pytest test/transformers/test_monkey_patch.py -q
# 66 passed

PYTHONPATH=src python3 -m pytest \
  test/convergence/bf16/test_mini_models.py \
  test/convergence/bf16/test_mini_models_with_logits.py \
  test/convergence/fp32/test_mini_models.py \
  test/convergence/fp32/test_mini_models_with_logits.py \
  -k deepseek_v3 -xvs
# 4 passed