Fix last-element indexing mixed with slices or scalar indices

[林SO (Linxiushen)]

Expressions such as x[-1, :2], x[:2, -1], and x[-1, -1] currently lower the last-element index to an empty Slice(-1, 0) before squeezing that axis. This raises an error in both eager evaluation and exported ONNX models

Use the axis end as the stop for scalar -1: the known dimension length in eager mode and the converter's existing maximum-int64 bound for exported models. Add public @script regressions comparing values and shapes with NumPy across two symbolic input shapes, including mixed scalar indices, reverse slicing, and a positive-index control

Validation:

• New regression matrix on unchanged main: 20 failed, 4 controls passed
• New regressions plus existing tensor, argument and evaluator tests: 58 passed using Python 3.12, ONNX 1.23 and CPU ONNX Runtime 1.30
• Repository-configured lintrunner -a on all three changed files and git diff --check: passed

The full repository and Torch/GPU suites were not run. This change addresses last-element indexing mixed with slices or scalar indices; it does not address general advanced-index handling