Direct FFT method: resolve image boundary condition

[antonysigma]

When FFT's target width and height are larger than the image size, pad the space with zeros. Also, if shiftx and shifty are non-zeros, resolve the W-by-H tile with cyclic boundary condition.

Typical use-case:

# Note the size difference between the input image and the blur kernel
input_image = np.empty((1024, 1024), dtype=np.float32, order='F')
blur_kernel = np.ones((3,3)) / 9.0
alpha = 1e-3

u = Variable((1024, 1024))
problem = Problem(
  sum_squares( conv(blur_kernel, u) - input_image ) + alpha * group_norm1( grad(u) ),
  implem='halide',
)

problem.solve()