[Feature]: hipTensor rocKE backend #14491

Suggestion Description

hipTensor runs on Composable Kernel (CK) instances. It would be desirable to add a rocKE backend to hipTensor next to CK, without changing the public API:

1. Backend interface:
   • an internal, any-rank problem descriptor and ISolution / IBackend / BackendRegistry;
   • the backend is selected at plan creation for every operation (contraction, permutation,
     elementwise binary/trinary, reduction);
   • plan cache entries are kept per backend.
2. CK stays the default with unchanged behavior. rocKE is opt-in:
   • -DHIPTENSOR_ENABLE_ROCKE_BACKEND=ON at build time;
   • HIPTENSOR_BACKEND=auto|ck|rocke at run time;
   • fallback to CK for problems rocKE declines (HIPTENSOR_BACKEND_FALLBACK).
3. rocKE kernels:
   • contractions in f16, bf16, f32, f64, complex f32 and complex f64 (scale and bilinear,
     unary operators on operands, conjugation), plus permutation, elementwise operations and
     reductions;
   • shape-generic kernels compiled at build time into a pack per GPU target, loaded and
     launched from C++ with no Python at run time.
4. Kernel selection at plan creation:
   • capability checks, then rocKE's tuned route table, then a selection model evaluated in
     C++;
   • optional on-device autotune of the top candidates;
   • the selected kernel's workspace reported at plan creation.

Operating System

No response

GPU

MI350, MI300

ROCm Component

hipTensor, rocKE