Add a fast path for batch-norm CPU inference. (#19152)

Add a fast path for batch-norm CPU inference. (#19152)

Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/19152

Adding a fast path for batch-norm CPU inference when all tensors are contiguous.
* Leverage vectorization through smiple loops.
* Folding linear terms before computation.
* For resnext-101, this version gets 18.95 times faster.
* Add a microbenchmark:
* (buck build mode/opt -c python.package_style=inplace --show-output //caffe2/benchmarks/operator_benchmark:batchnorm_benchmark) && \
(OMP_NUM_THREADS=1 MKL_NUM_THREADS=1 buck-out/gen/caffe2/benchmarks/operator_benchmark/batchnorm_benchmark#binary.par)
* batch_norm: data shape: [1, 256, 3136], bandwidth: 22.26 GB/s
* batch_norm: data shape: [1, 65536, 1], bandwidth: 5.57 GB/s
* batch_norm: data shape: [128, 2048, 1], bandwidth: 18.21 GB/s

Reviewed By: soumith, BIT-silence

Differential Revision: D14889728

fbshipit-source-id: 20c9e567e38ff7dbb9097873b85160eca2b0a795