dialects: dmp: make strategies attributes and carry them in IR

[PapyChacal]

As of now, the decomposition strategies are dataclasses and the information to construct them is expected to match in multiple bits of the pipeline (actually fused in distribute-stencil, which happens to do shape inference too, but I'm trying to decouple things here)

So instead, define them as attributes, and make the initial distribution pass pop that attribute on the swaps. The swap shape inference then do not need extra information to carry on.

Teaser: I want this part to be just collaborating shape inference as in:

• transformations: Implement shape-inference pass #3047

[codecov]

Codecov Report

All modified and coverable lines are covered by tests ✅

Project coverage is 89.88%. Comparing base (f591b3b) to head (3b8dd64).
Report is 1 commits behind head on main.

Additional details and impacted files

@@            Coverage Diff             @@
##             main    #3050      +/-   ##
==========================================
+ Coverage   89.87%   89.88%   +0.01%     
==========================================
  Files         416      415       -1     
  Lines       52441    52426      -15     
  Branches     8115     8111       -4     
==========================================
- Hits        47130    47123       -7     
+ Misses       4006     3999       -7     
+ Partials     1305     1304       -1     

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