[Polygraphy] Add clear error for mismatched ONNX Slice input lengths (#4470)

[njh1205]

Summary

TensorRT's ONNX parser rejects Slice ops whose starts/ends/axes/steps have inconsistent lengths with a low-level assertion (e.g. "Assertion failed: (starts.size() == axes.size())", see #4470), which is hard to diagnose. Since the TRT ONNX parser source is closed in the OSS release, this fix lands at the Polygraphy tooling layer: NetworkFromOnnx / NetworkFromOnnxPath now validate Slice input lengths before invoking the parser and raise a clear, actionable error naming the offending node.

Changes

• polygraphy/backend/trt/util.py: add check_onnx_slice_input_lengths() (resolves initializer / Constant-node / static-shape inputs).
• polygraphy/backend/trt/loader.py: run the check in both ONNX import paths.
• tests/backend/trt/test_onnx_slice_check.py: CPU-runnable regression tests (no TensorRT required).
• CHANGELOG.md: Unreleased entry.

Validation

• pytest tests/backend/trt/test_onnx_slice_check.py -> 7 passed (CPU).
• Before/after: the repro model from The onnx parser failed to parse a valid model: Slice (importSlice): INVALID_NODE: Assertion failed: (starts.size() == axes.size()) #4470 now reports 'Slice node ... has mismatched input lengths (starts=1, ends=2, axes=2)' instead of TRT's raw assertion.
• Formatted with black (Polygraphy pre-commit).

Fixes #4470 (tooling-layer solution; the parser itself is closed-source).

• End-to-end validated locally on NVIDIA L4 (driver 550) with TensorRT 8.5.3 Python bindings: the The onnx parser failed to parse a valid model: Slice (importSlice): INVALID_NODE: Assertion failed: (starts.size() == axes.size()) #4470 repro now surfaces the friendly error through the real TRT parser path (NetworkFromOnnxBytes), and a valid Slice model still imports cleanly (network parsed OK).

[pranavm-nvidia]

Let's fix it in the ONNX parser: https://github.com/onnx/onnx-tensorrt/blob/11.3-GA/onnxOpImporters.cpp#L6406. I think the fix would be to change:

        axes = nbInputs > 3 && !inputs.at(3).isNullTensor() ? ShapeTensor(ctx, inputs.at(3))
                                                            : iotaShapeVector(dims.size());

to:

        axes = nbInputs > 3 && !inputs.at(3).isNullTensor() ? ShapeTensor(ctx, inputs.at(3))
                                                            : iotaShapeVector(starts.size());

(which is already the behavior of the else branch).

[njh1205]

Parser-level fix now in place: onnx/onnx-tensorrt#1064 (11.3-GA) and #1065 (8.5-GA). See those PRs for the one-line parser fix and the regression test; closing this Polygraphy pre-check workaround in favor of the parser fix.