feat(kimi-k3): add expert LoRA and tune full SFT

[HuiyingLi]

What does this PR do ?

Adds a production-scale Kimi K3 ordinary-LoRA SFT example and updates the existing full-FT SFT recipe to use HybridEP with reshard-after-forward.

Changelog

• Add examples/llm_finetune/kimi/k3_hellaswag_lora.yaml with PP4/DP32/EP32, HybridEP, torch_mm experts, selective activation checkpointing, and reshard-after-forward.
• Apply ordinary rank-32 LoRA to all supported linear modules and enable moe_rank_scaling, which gives the 16-active-expert K3 MoE modules rank 2.
• Save a consolidated safetensors adapter at the final step.
• Update examples/llm_finetune/kimi/k3_hellaswag.yaml full FT from the torch dispatcher to HybridEP and enable reshard-after-forward.
• Add focused CPU recipe-contract tests for both the full-FT and expert-LoRA configurations.

Before your PR is "Ready for review"

Pre checks:

• Read and followed the contributor guidelines.
• Added the necessary focused tests.
• Added the runnable example recipe and inline topology documentation.

Validation

• pytest -q tests/unit_tests/recipes/llm/test_k3_lora_recipe.py (2 passed after the full-FT HybridEP/reshard update)
• pytest -q tests/unit_tests/recipes/llm/test_k3_lora_recipe.py tests/unit_tests/models/kimi_k3/test_state_dict_adapter.py tests/unit_tests/moe/test_moe_lora_state_dict.py tests/unit_tests/_transformers/test_recipe_doc_coverage.py (39 passed, 3 skipped before the full-FT config-only update)
• python tests/ci_tests/utils/validate_new_recipe_ci.py examples/llm_finetune/kimi/k3_hellaswag_lora.yaml
• ruff format .
• ruff check --fix .
• Four-GPU, four-layer K3 smoke: two train steps, validation, and consolidated adapter save passed with HybridEP + torch_mm + selective AC + resharding. The adapter contained all routed-expert LoRA tensors at expert rank 2.

Additional Information

Based on main after HybridEP token-count equalization in #3641. K3 expert-LoRA HF adapter key conversion is already covered by #3435. The full-FT HybridEP/reshard update is covered here by a config-contract test; it has not yet been rerun as a separate GPU job in this PR.

[copy-pr-bot]

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.