refactor(native): Split nn.cu into modular files (aligned with #131)

[m96-chan]

Problem

native/ops/nn/nn.cu is 2673 lines with mixed concerns:

• Activations (GELU, SiLU, Sigmoid, Tanh)
• Normalization (LayerNorm, RMSNorm)
• Attention (SDPA)
• Position encoding (RoPE)
• Tensor ops (Transpose, Bias Add)
• Linear layer

This should be split to match the binding structure in #131.

Current Structure

native/ops/nn/
├── nn.cu                    (2673 lines - everything)
├── attention_kernels.cuh
├── elementwise_kernels.cuh
├── flash_attention.cuh
├── memory_kernels.cuh
└── norm_kernels.cuh

Proposed Structure

native/ops/nn/
├── activation/
│   ├── gelu.cu
│   ├── silu.cu
│   ├── sigmoid.cu
│   ├── tanh.cu
│   └── relu.cu
├── norm/
│   ├── layernorm.cu
│   └── rmsnorm.cu
├── attention/
│   ├── sdpa_causal.cu
│   ├── sdpa_fixed_cache.cu
│   └── flash_attention.cuh
├── rope/
│   └── rope_inplace.cu
├── linear/
│   └── linear_bias.cu
└── common/
    ├── attention_kernels.cuh
    ├── elementwise_kernels.cuh
    ├── memory_kernels.cuh
    └── norm_kernels.cuh

Contents Mapping

Current Section	Lines (est.)	New Location
GELU	~50	activation/gelu.cu
SiLU	~60	activation/silu.cu
Sigmoid	~60	activation/sigmoid.cu
Tanh	~60	activation/tanh.cu
LayerNorm	~150	norm/layernorm.cu
RMSNorm	~100	norm/rmsnorm.cu
SDPA Causal	~400	attention/sdpa_causal.cu
SDPA Fixed Cache	~300	attention/sdpa_fixed_cache.cu
RoPE	~200	rope/rope_inplace.cu
Linear + Bias	~100	linear/linear_bias.cu
Transpose (3D/4D)	~200	Move to tensor/
Split QKV	~100	Move to tensor/

Notes

• Transpose operations should move to native/ops/tensor/
• Split QKV should move to native/ops/tensor/
• Each file should be <300 lines
• Shared kernels stay in common/

Benefits

• Faster incremental builds
• Easier to find and modify specific operations
• Aligned with binding structure (refactor(native): Split ops_bindings.cpp into modular files #131)
• Clear separation of concerns

Related

• refactor(native): Split ops_bindings.cpp into modular files #131 (ops_bindings.cpp split) - Should be done together
• refactor(api): Standardize GEMM/GEMV Python binding naming convention #130 (naming convention) - Align function names