kernel_tuning_net gtest is fragile due to violating blackbox

[Migrated from original issue] ROCm/MIOpen#3708

Original issue author: @randyspauldingamd

Test failure observed after updating CK:

[2025-05-01T10:36:12.397Z] /home/jenkins/workspace/MLLIBS_MIOpen_PR-3707/test/gtest/kernel_tuning_net.cpp:188: Failure
[2025-05-01T10:36:12.397Z] Expected equality of these values:
[2025-05-01T10:36:12.397Z]   perf_config.ToString()
[2025-05-01T10:36:12.397Z]     Which is: "DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1<128, 128, 32, 32, 8, 8, Default, 32, 32, 2, 1, 8, 8, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v1>"
[2025-05-01T10:36:12.397Z]   test_case.expected_config
[2025-05-01T10:36:12.397Z]     Which is: "DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1<128, 128, 32, 32, 8, 8, Default, 32, 32, 2, 1, 8, 8, 1, 1>"

@msaudulhassan reports that this is expected behavior when CK and the ktn models get out of sync. This makes the test inherently fragile, and it should be rewritten to use mocks and canned "gold" model files to achieve blackbox. In other words, a proper test shall test only the code itself, and not involve the actual version of CK nor model files.

The test shall be disabled for now. The priority and timeline are uncertain at this time, so if anyone can suggest an acceptable workaround, please do!

Start with test/gtest/kernel_tuning_net.cpp, especially the failure at line 188, and review how CK and ktn model files are used. Determine the mock and canned “gold” model boundaries described in the issue; done means the test checks its own behavior without depending on the installed CK version or live model files, with the current fragile test disabled until then.