Refactor LLaMA 4 kernels to modular nn/ structure

[m96-chan]

Problem

The current LLaMA 4 CUDA kernel implementation is monolithic:

native/ops/nn/
├── llama4/           # ← All kernels bundled together
└── llama4_kernels.cuh

This violates the modular architecture defined in CLAUDE.md where NN operations should be separated by function:

native/ops/nn/
├── activation/   # GELU, SiLU, etc.
├── attention/    # SDPA
├── norm/         # RMSNorm, LayerNorm
└── rope/         # RoPE

Impact

• LLaMA 4 specific kernels cannot be reused by other models
• Violates the principle of modular, composable operations
• Makes testing individual components harder

Proposed Solution

1. Extract common operations from llama4/ into their respective directories
2. Keep only LLaMA 4 specific logic (if any) in llama4/
3. Update bindings to use modular kernel paths

Related

• Added in commit 5fcf3c3 with note about needing refactor