Fix Polygraphy to-input multi-iteration merge

[Lingavasan]

Summary

• ensure polygraphy data to-input creates a distinct OrderedDict for each padded iteration
• add a regression test that preserves separate values for multiple input iterations

Context

This addresses the polygraphy data to-input aliasing bug described in #4607. The broader debug reduce multi-iteration limitation noted in that issue is not changed here.

Validation

• Exercised the actual ToInput().run_impl(...) path with two saved input iterations and verified both are preserved
• Reproduced the old [OrderedDict()] * n aliasing behavior and verified the new list-comprehension behavior creates distinct dictionaries
• PYTHONPATH=tools/Polygraphy /private/tmp/tensorrt-polygraphy-tests-py311/bin/python -m py_compile tools/Polygraphy/polygraphy/tools/data/subtool/to_input.py tools/Polygraphy/tests/tools/test_data.py
• git show --check --stat HEAD

Note: I could not run the official pytest tools/Polygraphy/tests/tools/test_data.py suite locally because tools/Polygraphy/tests/tools/conftest.py imports tensorrt during collection. TensorRT wheels are not available for this local macOS arm64 environment (Darwin 25.3.0, no nvidia-smi), so collection fails before reaching the test.

[Lingavasan]

Hi @wenbingl @akhilg-nv, just checking in when you have a chance.

This PR fixes the Polygraphy data to-input multi-iteration aliasing issue from #4607. The change is small and includes a regression test. Please let me know if you would prefer a different approach.

[Lingavasan]

Hi @wenbingl @akhilg-nv, following up on this PR when you have a chance.

It fixes the Polygraphy data to-input multi-iteration aliasing issue from #4607 with a small code change and regression test. DCO is passing and the PR is mergeable.

Please let me know if you would prefer any changes to the approach.

[Lingavasan]

Hi @wenbingl @akhilg-nv, following up when you have a chance.

This PR fixes the Polygraphy data to-input multi-iteration aliasing issue from #4607 with a small code change and regression test. The PR is still open and mergeable from my latest check.

Please let me know if you would prefer any changes to the approach.