perf(cuda): StreamTensor ops rewritten — 30+ element-wise ops GPU-native
stream_ops.rs: ZERO host-roundtrip in ANY element-wise operation.
Every gpu_* function now delegates to ElementwiseKernels CUDA kernels.
Rewrote: gpu_add, gpu_sub, gpu_mul, gpu_div, gpu_sigmoid, gpu_tanh,
gpu_silu, gpu_relu, gpu_elu, gpu_exp, gpu_log, gpu_abs, gpu_neg,
gpu_sqrt, gpu_sqr, gpu_sin, gpu_cos, gpu_floor, gpu_clamp, gpu_scale,
gpu_affine, gpu_recip, gpu_scalar_sub, gpu_add_scalar, gpu_minimum,
gpu_max_scalar, gpu_transpose, gpu_softmax, gpu_layer_norm, gpu_mean_all
Memory ops: gpu_cat_dim0/dim1, gpu_narrow_2d, gpu_select_dim1,
gpu_stack_2d — all DtoD async copies, zero host involvement.
elementwise.rs: 10 new CUDA kernel ops — sin, cos, sqrt, silu, elu,
recip, sigmoid, tanh, min, max. Broadcast row/col extended with add/sub.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
5f3c44af2b