MLAS: Add MlasComputeSoftmax/MlasComputeExp

[tracysh]

Description: This adds optimized routines to compute softmax/logsoftmax/exp.

Motivation and Context
This change adds:

1. MlasComputeExp function to vectorize exp() over a buffer used for the Exp op. The previous implementation used Eigen but was stuck the instruction set used to build ORT. The new version supports SSE2/AVX2/AVX512F (as well as NEON/VSX).
2. MlasComputeSoftmax function to vectorize softmax()/logsoftmax() for a NxD buffer. This is now used for the Softmax/LogSoftmax/Attention ops. This is optimized for the same platforms as MlasComputeExp.

This also restores threading to MlasComputeSoftmax for 1.3.

For an internal customer model, the changes improved a model from 134ms to 124ms per inference.

[skottmckay]

[yufenglee]