feat(ml-core): complete GpuTensor/GpuLinear/GpuVarStore API surface
GpuTensor: Clone derive (ref-counted CudaSlice), dims/dim/dims2/dims3,
from_vec/from_slice aliases, detach (no-op clone), powf/sqr/floor,
to_vec0/to_vec1, backward (no-op stub), relu/sum_all/flatten_all,
transpose/clamp/expand/broadcast_as/index_select. All host-side with
TODO: CUDA kernel markers.
GpuLinear: new(in, out, stream) standalone constructor,
weight_to_vec/bias_to_vec for checkpoint only.
GpuVarStore: cuda_stream(), data(), with_device(&MlDevice).
MlDevice: cuda_if_available(), new_cuda(), device() identity.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
e8dd460898