onnxruntime

Include algorithm selection exposed by ROCBLAS extensions API in GEMM autotuning (#13831)

### Description
Extend GEMM autotuning by including algorithms exposed by a ROCBLAS
extension API.

### Motivation and Context
Based on our request, the ROCm team has implemented extension APIs in
ROCBLAS that provides a list of application GEMM
algorithms/implementations for a given input size, along with an API
that actually performs the GEMM using the specified
implementation/algorithm. We have observed that the ROCBLAS
algorithm/implementation selection logic does not always pick the
optimal. This PR uses the extension APIs to integrate the exposed
ROCBLAS algorithms/implementations into the autotuning framework.

The feature is disabled by default (the ROCBlas extension APIs are
slated to be released with ROCm 5.5, and are not yet generally
available). To enable: build with `--cmake-extra-defines
USE_ROCBLAS_EXTENSION_API=1 CMAKE_HIP_FLAGS=-DUSE_ROCBLAS_EXTENSION_API`
and then enable tuning in the provider options.

Co-authored-by: Abhishek Udupa <abhishek.udupa@microsoft.com>