Add rotating tensor, icache flush, and bias to GEMM tuning script

[scxiao]

This PR is to add three new features to the gemm tuning script tune_gemm.py

• added --rotating_tensor <value> to use different memory blocks in each iteration when running gemm performance, size in MB, default is 0MB, so no rotating tensor by default.
• added --icache_flush to flush icache in each iteration when running GEMM performance. Rotating tensor and icache flush are to make perf numbers closer to real applications.
• added --bias_vector to turn on bias add to output matrix. The bias vector contains the same number of elements as the number of rows of the output matrix. Each element is added to all elements of the corresponding row of the output matrix.

[scxiao]

Hi all, I change the icache flush processing as: since python-hip is needed to run the icache-flush kernel, if the python-hip module is not installed and icache-flush is enabled, it will print a message mentioned no python-hip, so icache-flush is disabled. The message is as:

************************************************************************************************
  `icache-flush` is disabled for this run.
  `icache-flush` needs python-hip module, which is unavailable.
  python-hip module can be installed as:
      `python3 -m pip install -i https://test.pypi.org/simple hip-python~=$rocm_version`
************************************************************************************************

In this way, the script can still run to completion without icache-flush, or user can install python-hip as in the message.

Also, if no further comments, could anyone help approve this PR, so I can get it merged.

[zhanglx13]

Also, can you change the title and description of this PR to include support for icache_flush and bias. And I think this is a big change so I'd suggest we bump the version of tune_gemm to be v3.2 (I'm suggesting without any experience about versioning, we can discuss later about it) and can you update the added features in the README?