refactor(ops): Split nn.py to match native nn/ structure

[m96-chan]

Problem

src/pygpukit/ops/nn.py is 885 lines and should match the native nn/ structure from #133.

Current State

src/pygpukit/ops/
└── nn.py  (885 lines - all NN ops)

Proposed Structure (aligned with #133)

src/pygpukit/ops/nn/
├── __init__.py          (exports)
├── activation.py        (gelu, silu, relu, sigmoid, tanh)
├── norm.py              (layernorm, rmsnorm)
├── attention.py         (sdpa_causal, sdpa_causal_fixed_cache)
└── rope.py              (rope_inplace, rope_inplace_f32table)

Benefits

• Mirrors native/ops/nn/ structure
• Each category < 200 lines
• Easier to maintain Python/C++ correspondence

Related

• refactor(native): Split nn.cu into modular files (aligned with #131) #133 (native nn.cu split)
• refactor(native): Split ops_bindings.cpp into modular files #131 (binding split)

[m96-chan]

Completed in PR #168