Create initializers not constant nodes in constant folding pass

[titaiwangms]

Partially From #2598

This provides a better optimized graph after constant folding in terms of the number of nodes, which is better for debugging.

[codecov]

Codecov Report

❌ Patch coverage is 91.89189% with 3 lines in your changes missing coverage. Please review.
✅ Project coverage is 70.24%. Comparing base (45b5189) to head (c410cbc).

Files with missing lines	Patch %	Lines
onnxscript/optimizer/_constant_folding.py	80.00%	1 Missing and 2 partials ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##             main    #2650      +/-   ##
==========================================
- Coverage   70.41%   70.24%   -0.17%     
==========================================
  Files         224      224              
  Lines       26600    26605       +5     
  Branches     2645     2645              
==========================================
- Hits        18730    18689      -41     
- Misses       6945     6991      +46     
  Partials      925      925              

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

[titaiwangms]

Need onnx/ir-py#227

[titaiwangms]

This might be considered bc breaking in terms of graph changes.

[justinchuby]

Thanks!

[justinchuby]

Just making sure: is the replaced initializer immediately freed (if there is no users)? We need to do that to avoid memory usage jumps.

[titaiwangms]

Just making sure: is the replaced initializer immediately freed (if there is no users)? We need to do that to avoid memory usage jumps.

I will check.