Use a stable lowering for ReduceLogSumExp

[peter941221]

Refs NVIDIA/TensorRT#4772

Problem
The ReduceLogSumExp importer exponentiates the input before reduction. Large finite inputs overflow in that first exp, so TensorRT returns inf where ONNX Runtime stays finite.

Change
Use the stable form max(x) + log(sum(exp(x - max(x)))) for ReduceLogSumExp lowering.
Reduce the max over the same axes as the sum.
Unsqueeze the reduced max when keepdims=0 so the subtraction still broadcasts correctly.
Preserve noop_with_empty_axes=1 when opset 18 passes an empty axes input.
Extend the regression coverage for large finite inputs, keepdims=0 with axes input, and noop_with_empty_axes=1 with an empty axes input.

Validation
I reproduced NVIDIA/TensorRT#4772 with TensorRT 10.16.1.11.
Input: [250.0, 248.0, 255.0, 251.0]
ONNX Runtime: 255.025634765625
TensorRT before: inf
TensorRT after patched parser: 255.025634765625

I rebuilt the parser against the 10.16 source tree and validated it by temporarily swapping the parser library in a TensorRT 10.16.1.11 Python environment.
python -m unittest -v onnx_backend_test.TensorRTCustomReduceLogSumExpTest