Matrix Multiplication Improvements

[andrej]

• work around object fifo bug #1547
• in host verification code, accumulate in float instead of bf16 -- this allows verification to pass even for bigger problem sizes, where before the accumulation errors got too big
• if problem size is beyond threshold, stochastically verify instead of exhaustively (exhaustively verifying takes way too long, which is probably why the failed verification for larger problem sizes went undetected before) -- if this is merged, we can close #1134 which contains a similar change
• add assertions and comments
• improve verification output
• add workaround comment for excessive program size (CDO generation error [AIE ERROR] _XAie_LoadProgMemSection():230: Overflow of program memory e.g. for size M=768 K=768 N=256 if fifo_depth>1)

cc @jgmelber

[andrej]

• make use of changes in #1538 to simplify code (strides) for whole_array design
• allow changing inner tile size (small m, k, n) for whole_array design
• replace complicated verficiation code with the naive implementation; the idea was that the other implementation would be faster, but it was not (significantly); swapped the naive version back in to have easier maintainability and confidence in the code

[andrej]

@jgmelber Kindly kick off the tests for this when you get a chance. I fixed a number of issues with the matmuls that slid by because the verification tolerance was a bit high. I reduced the tolerance, so let's see if the tests reveal more issues.

[andrej]

@jgmelber Could you please kick off the CI again? TIA

[jgmelber]

@andrej matrix_vector still needs to be fixed, the rest look okay

[andrej]

@jgmelber Hopefully this will pass now

[andrej]

@jgmelber Could you kick off the tests on this please?

[andrej]

Looks like tests pass but I forgot to format. Another try please @jgmelber ?