[transformers] Add tiled MLP patching via the transformers register_patch_mapping API

[kashif]

Liger already ships a memory-efficient tiled SwiGLU/GEGLU MLP, but there was no way to reach it through the normal patching path. You had to construct the module by hand, so the activation-memory win was hidden from apply_liger_kernel_to_* / use_liger_kernel=True.

This wires it in using the official transformers monkey patching API, with one mechanism rather than per-model edits:

• apply_liger_tiled_mlp(model=None, num_shards=None):
  • pre-load (model=None): registers the swap via transformers.monkey_patching.register_patch_mapping, so from_pretrained / from_config build the tiled MLP directly. Weights load by name since gate/up/down projections match.
  • instance (model=...): walks the model and patches in place every MLP whose class name is in the mapping, reusing existing weights. This covers the already-loaded case the official API cannot retrofit.
• _apply_liger_kernel_to_instance gains tiled_mlp / tiled_mlp_num_shards, so use_liger_kernel=True and TRL can opt in uniformly for every supported model.
• LIGER_TILED_MLP_PATCH_MAPPING lists the 15 models that share the SwiGLU or GEGLU layout (llama, mistral, ministral, qwen2/3, qwen2_vl, smollm3, mllama, pixtral, llama4, olmo2/3, exaone4 -> tiled SwiGLU; gemma/2/3 -> tiled GEGLU).

MoE (mixtral, qwen3_moe), fused gate_up (phi3), Gemma4 and Granite are intentionally excluded since they use bespoke layouts and need their own tiled variants.

Distributed note: tiled MLP is strictly opt-in. Its distributed behaviour is whatever the underlying LigerTiledMLPFunction provides, which targets DeepSpeed/ZeRO. FSDP/FSDP2 and plain DDP are not yet handled by the default tiled backward (see open PRs #1128 and #1125), so enabling tiled MLP under those backends can corrupt gradients today. This PR only does the wiring and does not change that kernel.

Tests cover the central instance path, the standalone instance patch, and the pre-load registry across both a silu and a gelu model. Full test_monkey_patch.py passes (65).

[kashif]

cc @vaibhavjindal