This PR

1. Implements short-time fourier transform add torch.stft and torch.fft #3775 with tests against Scipy.
2. Adds atan2 for cuda double & half tensors. Refactored atan2 to THCNumerics, Op to THCTensorMathPointwise.cuh.
3. Adds support for python only default init in the parsers. This allows us to write argument with a default value that can either cause ambiguity in c++ (e.g., Scalar p in norm) or have a type that doesn't allow default value None/NULL/nullptr (e.g., int64_t fft_size in stft).
   a. Changes norm to also use this in both Declarations.cwrap for torch.norm named arguments do not work as expected #1419 . When Tensor and Variable merge, it should fix the issues in torch.norm named arguments do not work as expected #1419 .
4. Implements three commonly used window functions (Hann, Hamming, Bartlett) with tests against Scipy

On 3: Now in python_variable_methods.cpp, if an argument has python only default value, it will be generated like the following. I added the comment for clarification purpose. It won't be in actual code.

static PyObject * THPVariable_stft(PyObject* self, PyObject* args, PyObject* kwargs)
{
  HANDLE_TH_ERRORS
  static PythonArgParser parser({
    "stft(int64_t frame_length, int64_t hop, int64_t fft_size=None, Tensor window=None, int64_t pad_end=0)",
  });
  auto& self_ = reinterpret_cast<THPVariable*>(self)->cdata;
  PyObject* parsed_args[6];
  auto r = parser.parse(args, kwargs, parsed_args);
  if (r.idx == 0) {
    Tensor & self = self_;
    int64_t frame_length = r.toInt64(0);
    int64_t hop = r.toInt64(1);
    const Tensor & window = r.tensor(3);
    int64_t pad_end = r.toInt64(4);
    // fft_size has python only default, so it will be generated last
    int64_t fft_size = r.toInt64WithDefault(2, frame_length);  
    return wrap(dispatch_stft(self, frame_length, hop, fft_size, window, pad_end));
  }
  Py_RETURN_NONE;
  END_HANDLE_TH_ERRORS
}