Add Kaldi Pitch feature

[mthrok]

This PR adds Kaldi Pitch Feature detailed in "A pitch extraction algorithm tuned for automatic speech recognition".

The interface is mostly same as compute-kaldi-pitch-feats CLI. Batch support is added via at::parallel_for function.

As the function binds the custom built libkaldi, it only supports CPU (and float32) at the moment.

About the custom built Kaldi

Since kaldi is a large library, I choose to use only a subset of it by adding custom build process. In addition to that, to reduce the dependencies and make the build process, I reused the BLAS package PyTorch uses. For this, I added custom interface of Kaldi's matrix libraries. Therefor the algorithm runs on torch::Tensor class. However, some parts of the algorithms requires direct access to memory, so the resulting function is not differentiable. Also the resulting code is very slow (60x) at the moment due to the overhead of slicing operation. (Kaldi's feature implementations work element-wise, while PyTorch operates faster when operations are vectorized.)

Supersedes #1063

[cpuhrsch]

If this is called in a tight-loop with small Tensors, you might fare better inlining the operation and using the underlying tensor data pointer instead of calling into torch::mm repeatedly.

Also note that torch::mm does parallelization, but at::parallel_for disables nested parallelism for OpenMP (to avoid oversubscription), which is the default threadpool for PyTorch.