fix: prevent initializer name collision in rewrite rule registration

[OYCN]

Summary

#2945

Fix a bug where multiple rewrite rule matches producing initializers with the same name would silently clobber each other, resulting in a dangling ir.Value reference and a toposort validation failure on serialization.

Problem

When two Reshape→Reshape chains share the same shape initializer (e.g. containing -1), the ReshapeReshape rule fires twice, each time calling op.initializer(ir.Tensor(..., name=shape.name)). Both produce an initializer with the same name but different resolved values.

The registration logic in _rewrite_rule.py had two independent for-loops:

# Loop 1: check for duplicates (dead code — `continue` only skips this loop)
for initializer in delta.new_initializers:
    if initializer.name in initializers:
        continue

# Loop 2: unconditionally overwrite
for initializer in delta.new_initializers:
    initializers[initializer.name] = initializer

The second loop always overwrote the dict entry, so the first registered Value lost its dict reference. NameFixPass would then rename it to ensure uniqueness, but never re-registered it as an initializer — leaving it as a dangling reference that fails onnx.checker.

Fix

Merge into a single loop. On name conflict with a different Value object, auto-generate a unique suffix (_1, _2, ...) and register under the new name. Since node inputs hold object references (not name strings), the rename propagates correctly on serialization.

Test

Added test_reshape_reshape_shared_shape_no_clobbering covering the shared-shape scenario. Asserts that all shape Values are registered initializers and the serialized model passes onnx.checker.check_model(full_check=True).

Related

Also filed a separate issue/PR in onnx-ir for NameFixPass not registering renamed Values as initializers — that's an independent bug exposed by the same scenario.

[OYCN]

@microsoft-github-policy-service agree