Multi class LR behaves differently on .NetCore 3.0

[yaeldekel]

Four tests are failing on the .NetCore 3.0 builds:
EnsemblesMultiAveragerTest
EnsemblesMultiClassBootstrapSelectorTest
EnsemblesMultiVotingCombinerTest
EnsemblesMultiStackCombinerTest

The difference in the baselines is something like this:
L1 regularization selected 13 of 15 weights. vs.
L1 regularization selected 11 of 15 weights.

[frank-dong-ms-zz]

Investigating now, caused by different version of CPUMath gives slightly different result

[frank-dong-ms-zz]

The difference comes from CPUMath using different instruction set:

1. net framework and net core 2.1 uses CpuMathUtils.netstandard that uses SSE instruction set;
2. net core 3.1 uses CpuMathUtils.netcoreapp that uses AVX, SSE or direct floating point calculation depending on hardward avaibility.
   AVX and SSE generates slightly different result due to nature of floating point math.
   Use below issue to track: support AVX instruction set at CPUMath native #5044

[frank-dong-ms-zz]

created separate baseline for netcore app 31