[E2E Model Support] Add TileGym kernel integration for LFM2 MoE

[iamanishx]

Overview

This issue tracks the addition of end-to-end TileGym kernel support for Liquid AI's LFM2 MoE model family, starting with: LiquidAI/LFM2-8B-A1B
LFM2-8B-A1B is a hybrid MoE model with 8.3B total parameters and 1.5B active parameters. It combines full-attention blocks, convolution-style blocks, GQA attention, RoPE, RMSNorm-style normalization, SwiGLU-style expert FFNs, and top-k MoE routing.

Planned Steps

• Add apply_tilegym_kernel_to_lfm2_moe in monkey_patch.py
• Register lfm2_moe in MODEL_TYPE_TO_APPLY_TILEGYM_FN
• Patch compatible common kernels where applicable:
  • RoPE
  • RMSNorm
  • attention path
  • SwiGLU / expert MLP path
• Add an LFM2 MoE wrapper that reuses TileGym fused_moe
• Verify expert weight layout and routing semantics:
  • gate/up projection order
  • down projection layout
  • top-k routing weights
  • norm_topk_prob
  • routed_scaling_factor
  • expert bias handling
• Add an E2E inference / benchmark script for LiquidAI/LFM2-8B-A1B

Questions

• Is LiquidAI/LFM2-8B-A1B an acceptable validation target for a 5090-friendly E2E model integration?
• Are there existing TileGym conventions for hybrid architectures with non-attention convolution blocks that this integration should follow?