address issues raised in PR #2154

[bghimireamd]

• use struct to return data
• see if passing of unique_ptr is really necesary
• remove item added to target_link_library that was part of workaround

This issue is created to document and address issue seen in PR #2154 .

[atamazov]

@bghimireamd [Notice] Thanks for creating the descriptive post-merge issue for your PR. I dream of a bright future where all developers do the same 🌴 🌞

[atamazov]

• [x] #2239
  • Via #2241

[atamazov]

• [ ] https://github.com/ROCmSoftwarePlatform/MIOpen/pull/2154/files#r1251332706
  • Problem: We have no tuning test for fused tunable solvers (e.g. ConvBiasActivAsm1x1U)
    • Note: The "parent" of ConvBiasActivAsm1x1U is ConvAsm1x1U. It is for gfx9 only AND deprecated (listed in #2080 and thus by default applicable only for MI50/60 and MI100/200)

[atamazov]

• [ ] https://github.com/ROCmSoftwarePlatform/MIOpen/pull/2154/files#r1252350092
  • Problem: If tuning is enabled in the test (see previous comment), then it won't trigger reliably. And when triggered it will most likely take too long.
    • NOTE: To date, MIOPEN_DEBUG_TUNING_ITERATIONS_MAX is in develop for all solvers except Legacy OCL ones. Support for legacy solvers resides in #1911. You may wish to cherry-pick it from there into your development branch when fixing this comment.

[ppanchad-amd]

@bghimireamd Is this fixed with latest ROCm 6.1.0? If so, please close. Thanks!