Fix: Stable implementation for ReduceLogSumExp (ONNX frontend)

[sshekhar563]

Summary #32839

This patch fixes numerical instability in the ONNX frontend’s ReduceLogSumExp operator.
The previous implementation used a naïve log(sum(exp(x))) formulation, which overflowed to Inf for values near log(MAX_FLOAT32) (≈ 88.72). This caused incorrect results in the OpenVINO EP of ONNX Runtime, especially for float16 and float32 models.

Fix

Implemented a numerically stable LogSumExp computation:

k = ReduceMax(x)
lse = k + log( ReduceSum( exp(x - k) ) )

This matches the behavior of:

• ONNX Runtime CPU EP
• PyTorch and NumPy stable LogSumExp
• OpenVINO PyTorch frontend implementation (log.cpp)

The new implementation is applied to:

• Opset 1–12 (opset_1::reduce_log_sum_exp)
• Opset 13–17 (opset_13::reduce_log_sum_exp)
• Opset 18+ (opset_18::reduce_log_sum_exp)

Motivation

Fixes incorrect Inf outputs for values ≥ 88.7 when using OpenVINOExecutionProvider.
The issue was originally reported here: [link to GitHub issue].

Validation

Tested using provided reproduction script from the issue:

• Matches ONNX Runtime CPU EP for all tested values
• No overflow observed for large positive inputs
• Behavior consistent across float16/float32 models

Notes

This fix aligns ONNX frontend behavior with the already-correct PyTorch frontend implementation.

[cknd]

could someone of the team approve the CI workflows to move this forward?