Restructure code gen for FBGEMMFP16 (pytorch#236)

Summary:
Pull Request resolved: pytorch#236

1) Merge 1st iteration with data load and compute, reduce pressure from load port. Execute load and fma in parallel
2) Mask latency of Fp16->Fp32 conversion. Start conversion of B values required by the next iteration. !!! If matrix B size is exactly a page will generate segfault
3) Add per-matrix prefetch distance control

Codegen built with:
g++ --std=c++11 fbgemm/src/codegen_fp16fp32.cc -o code_gen
./code_gen

Reviewed By: dskhudia

Differential Revision: D19222915

fbshipit-source-id: 3cef33e99f636c6355245a4618b03787a06e41fa