Fix SBGEMM and SHGEMM test tolerances so that unit tests pass

[ChipKerchner]

SBGEMM and SHGEMM should have an acceptable tolerances so that unit tests pass.

SBGEMM has 7 bits of fraction and SHGEMM has 10.

Unit tests are failing (I know RVV is) because they are off by a couple of epsilon differences.

Please change so that we don't see these failures for each patch updated.

[ChipKerchner]

BGEMM might have the same issue.

[martin-frbg]

Not sure what's going on there - especially why these failures only pop up with the new SGEMM kernel in your draft PR. The BGEMM I understand could be caused by the SGEMM reference results shifting (but that would also mean shifting compared to other architectures ??)
The (excessive, I agree) output from the failing SHGEMM test seems to suggest that all of a sudden the difference between the SHGEMM result matrix elements and their counterparts from a naive multiplication has grown to give less than one decimal of accuracy

[ChipKerchner]

My previous change 5640 did not show these failures. I can not find changes that are RVV specific past that (maybe I overlooked something). My draft only affects SGEMM so it should NOT have any effect on SHGEMM. Does SHGEMM compare results to SGEMM?

[martin-frbg]

I see now that today's WebAssembly PR has the same excessive CC..DD output, but does not register the CI job as failed. Let me silence that stupid printf first...