fix: normalize cache key dtype to prevent initializer name collisions

[gramalingam]

When a Python int literal (e.g. 1) is used in both untyped positions (like a Gather index, where the ONNX schema type variable is unbound) and typed positions (like Add with an INT64 tensor), the constant cache created two entries: (1, None) and (1, INT64). Both generated the same initializer name 'const_1_i64' but as different ir.Value objects, causing register_initializer to raise ValueError.

Fix: before cache lookup, normalize dtype=None to the default ONNX dtype for the Python type (_PYTHON_TYPE_TO_DTYPE: int->INT64, float->FLOAT). This merges both entries into a single cache key and reuses the same ir.Value. Applied to both scalar and sequence (list/tuple) branches.

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 72.30%. Comparing base (13f265c) to head (e957e18).
⚠️ Report is 1 commits behind head on main.
✅ All tests successful. No failed tests found.

Additional details and impacted files

@@            Coverage Diff             @@
##             main    #2888      +/-   ##
==========================================
+ Coverage   72.24%   72.30%   +0.05%     
==========================================
  Files         241      241              
  Lines       29650    29678      +28     
  Branches     2914     2916       +2     
==========================================
+ Hits        21421    21459      +38     
+ Misses       7235     7227       -8     
+ Partials      994      992       -2     

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