Keep integer precision in rounded division

[Rio Yu (rioyu123)]

Summary

Integer torch.div(..., rounding_mode="floor") and rounding_mode="trunc" currently pass through float32 during export. Large integer quotients can consequently change value: for example, 50331653 / 3 with either rounding mode produces 16777218 instead of 16777217 in ONNX Runtime.

This change keeps rounded integer division in integers. trunc uses integer Div, and floor reuses the existing aten_floor_divide implementation, including its correction for negative operands. Floating-point inputs and the rounding_mode=None path are unchanged.

Fixes #3062.

Testing

• Added eight end-to-end export tests covering int32/int64, both rounding modes, tensor/scalar divisors, broadcasting, positive and negative operands, zero, exact multiples, and large quotients. Results are compared to PyTorch with zero tolerance. All eight fail before the fix and pass afterward.
• Existing division OpInfo coverage: 16 passed, 11 skipped, 135 subtests passed.
• Additional export/runtime smoke checks for int8, int16 and uint8 passed for both rounding modes.
• Full e2e_ops_tests.py: 135 passed, 1 skipped, 5 failed. Unmodified main has the same five failures (STFT value_info, boolean SDPA, and dynamic unbind), with 127 passed and 1 skipped.
• lintrunner -a and git diff --check passed.

Tested on Windows with Python 3.12, PyTorch 2.14.1 CPU and ONNX Runtime 1.31.0. TensorRT and model-quality benchmarks were not run.