Fix issue with fast_math in internal and upstream Onnxruntime MIGraphX EP

[TedThemistokleous]

It's come to my attention a patch in Onnxruntime MIGraphX EP wasn't upstreamed to our latest 6.1 and 6.2 branches.

Need to patch this + also ensure this is up streamed to mainline Microsoft/Onnxruntime

Patch in Question: ROCm/onnxruntime#25
Suspect this is the cause of this issue in JIRA: https://ontrack-internal.amd.com/browse/SWDEV-453466

Will need to also follow up with other perf issues where Driver vs EP is slightly faster: https://ontrack-internal.amd.com/browse/SWDEV-446822

Highly unlikely as I think the build used for this issue was after this batch was added and initial tests were done on 6.0.2 where this patch lives.

• Patch changes to 6.1_internal_testing
• Patch changes to rocm6.2_internal_testing
  - [ ] Patch changes to rocm/onnxruntime: main
• Upstream Patch Onnxruntime main
• Reconfirm with QA that issue is resolved for : https://ontrack-internal.amd.com/browse/SWDEV-453466
• Confirm if https://ontrack-internal.amd.com/browse/SWDEV-446822 has or doesn't have patch.

[TedThemistokleous]

https://ontrack-internal.amd.com/browse/SWDEV-446822 appears to have the fast_math batch applied, going by build numbers (115) on that ticket, and confirming with DevOps that the repo is cloned when using an internal testing branch to create the associated ROCm X.X files when we build the EP. It appears this is a different unrelated issue.

PR's to internal branches have been opened and awaiting review.

[TedThemistokleous]

Solves the SWB so closing this ticket out. Will hammer Onnxruntime upstream until we get this fix in as this is critical for GELU accuracy.