[codegen][gpu] Performance investigation of default convolution filter layout

[jerryyin]

Request description

This is a medium/low priority follow-up of #19701.

By the end of the ticket, we decide to use FHWC layout for filter by default in the preprocessing pipeline #19974. This is due to:

• For FHWC layout, the gemmK dimensions HWC will be collapsed as a single reduction dimension which is cleaner compared with HWFC layout
• FHWC is a more common choice among different framework and libraries
• FHWC and HWFC, in the limited tuning benchmarking, delivers around similar performance

However, the performance evaluation is not carried in a comprehensive manner and we may be able to adopt a better default layout in different scenarios. This ticket is used to track and further determine the right default layout for convolution in the preprocessing pipelines.

To finish this, we'd like a comprehensive study of the layout impact on convolution tuned performance and document our studies, alter the implementation of iree-preprocessing-convert-conv-filter-to-channels-last when necessary.

What component(s) does this issue relate to?

No response

Additional context

No response