transformations: (memref-streamify) don't streamify 0D memrefs

[superlopuh]

There's no point in streaming 0D memrefs, it costs more to set up the streaming registers. In the future we might want to canonicalize away the 0D memrefs on memref_stream operations to loads/stores and generic on scalars, which will be lowered to registers.

Note stacked PR.

[codecov]

Codecov Report

All modified and coverable lines are covered by tests ✅

Project coverage is 91.29%. Comparing base (6f59234) to head (28833de).
Report is 2 commits behind head on main.

Additional details and impacted files

@@           Coverage Diff           @@
##             main    #3677   +/-   ##
=======================================
  Coverage   91.29%   91.29%           
=======================================
  Files         466      466           
  Lines       58357    58357           
  Branches     5624     5624           
=======================================
  Hits        53278    53278           
  Misses       3629     3629           
  Partials     1450     1450           

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

[mamanain]

Don't you want to deal with this during linalg -> stream_memref lowering? Just replace it with load/op/store there.

[superlopuh]

What's the advantage of doing it when transforminglinalg to memref_stream? I would rather keep that pass as simple as possible, and do all the transformations/legalizations in downstream passes. We can transform the memref<f64> inputs to scalars in a separate canonicalization pass.

[mamanain]

LGTM