refactor(bindings): split ops_bindings.cpp into modular files (#131)

[m96-chan]

Summary

• Split monolithic ops_bindings.cpp (~3000 lines) into 39 organized binding files
• Created bindings_common.hpp with shared includes and forward declarations
• Organized by operation category with consistent naming convention
• ops_bindings.cpp now only ~77 lines (init function calls)

New Directory Structure

native/bindings/
├── bindings_common.hpp      # Common header
├── ops_bindings.cpp         # Main entry (calls init_* functions)
├── elementwise/             # binary, inplace, compare
├── unary/                   # math, trig
├── reduction/               # basic, argmax, softmax
├── tensor/                  # cast, transpose, reshape, repeat
├── embedding/               # lookup, kv_cache
├── nn/                      # activation, norm, attention, rope
├── gemm/                    # generic, fp8x*, nvf4x*, grouped, int
├── gemv/                    # generic, fp8x*, nvf4x*
├── sampling/                # basic, topk, seed
└── [other]                  # quantize, paged_attention, etc.

Test plan

• Build passes (SM120a, CUDA 13.1)
• 238 tests passing
• Pre-commit checks pass (ruff lint, ruff format, mypy)

Closes #131

🤖 Generated with Claude Code