Preserve Polygraphy cosine similarity under rescaling

[sylvesterkaczmarek]

Fixes #4868.

Scale each nonzero vector by its maximum absolute component before calculating cosine products and norms. This avoids overflow/underflow changing the comparison verdict for rescaled finite outputs. Preserve the zero-vector convention and non-finite-result rejection.

The scale is obtained from the extrema without the general absolute-value helper, whose PyTorch implementation first converts to float32. Integer inputs are promoted before scaling so their minimum representable value can be handled safely.

Add NumPy/PyTorch regressions for independently scaled identical, opposite and orthogonal vectors, plus empty, zero and signed-integer controls. The controls also check that inputs are unchanged. Update Polygraphy's changelog.

Validation

From tools/Polygraphy:

PYTHONPATH=. python -m pytest -q tests/comparator/test_compare.py   tests/comparator/test_struct.py tests/comparator/test_postprocess.py

283 passed, 4 existing skips on CPU. The 24 floating-point rescaling cases give 18 failures and 6 passes on unchanged upstream. Polygraphy's Black pre-commit hook and git diff --check pass.

No TensorRT engine, CUDA execution, or model accuracy benchmark was run. The tests exercise comparison arithmetic and verdicts using controlled arrays.