Fix TKW gemm tests

[Hardcode84]

• Use TKW create_vmfb_file option to get vmfb file directly
• Get scheduling params from the TK itself
• Disable scheduling for large K as compilation taking too long
• Support case when out_type != accumulator_type
• Update exec_args as code generated by TK expected out param to be present
• Gracefully handle TK errors instead of terminating entire script

[saienduri]

Thanks for the fixes Ivan! Is it possible to stick with generate TK mlir and then go through the normal compilation flow? It would be great if we can limit branching off from the general flow only when necessary. This allows us to scale and build an overall clean pipeline and helps also on visibility of how we are compiling (flags, commands, etc)

[Hardcode84]

Is it possible to stick with generate TK mlir and then go through the normal compilation flow?

I'm not sure if it's a good idea, TK can potentially add a custom flags to its default IREE pipeline and bypassing it means we are testing non-user facing version of the product.

[saienduri]

Ok that is fine. But, do you have an idea how we can make it clearly visible what flags are being used. Maybe in a comment or through logs. For this PR, this is ok though