Perf: Optimization for FactorizationMachine

[glebuk]

Benchmarking using VTune has found several bottlenecks in Factrization Machine training algorithm.

• Some phases training that are not parallelized. Consider adding parallel computation.
• Evaluate using AVX/AVX2 (C++ or C# instrinsics) in factorizationmachinenative.dll which currently implements C++ SSE code
• Consider optimizing the following hotspot

Function	Module	CPU Time
Microsoft::ML::Internal::Utilities::DoubleParser::TryParseCore	Microsoft.ML.Core.dll	27.390s
CalculateGradientAndUpdateNative	factorizationmachinenative.dll	22.609s
HelperImpl::FetchNextField	Microsoft.ML.Data.dll	13.826s
CalculateIntermediateVariablesNative	factorizationmachinenative.dll	12.201s
Microsoft::ML::Internal::Utilities::DoubleParser::TryParse	Microsoft.ML.Core.dll	9.666s

[frank-dong-ms-zz]

Done in below PR, close the issue now:
#3940