refactor(native): Split matmul.cu dispatcher into per-dtype files

[m96-chan]

Problem

native/ops/matmul/matmul.cu is 664 lines with dispatch logic for all dtypes.

Current State

native/ops/matmul/
└── matmul.cu  (664 lines - all dispatch)

Proposed Structure

native/ops/matmul/
├── dispatch.cu          (main dispatcher, ~100 lines)
├── dispatch_f32.cu      (FP32/TF32 dispatch)
├── dispatch_bf16.cu     (BF16 dispatch)
├── dispatch_fp8.cu      (FP8 dispatch)
└── dispatch_nvf4.cu     (NVF4 dispatch)

Benefits

• Faster incremental builds
• Clear dispatch path per dtype
• Easier to add new dtype support

Related

• refactor(native): Split ops_bindings.cpp into modular files #131 (binding split)
• refactor(ops): Split matmul.py into modular files #139 (Python matmul.py split)

[m96-chan]

Completed in PR #168