refactor(compute): split gpu_engine.go into focused files

Extract methods from gpu_engine.go (3521 -> 2245 lines) into three
focused files to improve navigability:

- gpu_engine_elementwise.go (400 lines): Add/Sub/Mul/Div, scalar ops,
  Exp/Log/Sin/Cos/Tanh/Pow, Sqrt/Rsqrt, fused RoPE/SwiGLU/RMSNorm,
  CosineSimilarity, HadamardTransform

- gpu_engine_reduction.go (221 lines): Sum, Softmax, ReduceSum/Max/Mean,
  GPUArgmax, GPUScaledSoftmax, GPUFusedSoftmaxVMul

- gpu_engine_memory.go (695 lines): Transpose, Zero/Zeros/Copy, Gather,
  ScatterAdd, Fill, Split/Concat/Repeat/RepeatInterleave, Reshape,
  OneHot, ConvertFP16ToF32

Zero behavioral changes. All method signatures identical. Build, vet,
and race-detector tests pass.

Author: dndungu