Fix dft fields memory leak

[smartalecH]

Fixes #1262

Fixes three issues currently found in the _get_dft_array routine.

1. Initializes the dimension object dim. 2D simulations don't set the third dimension by default, consequently whatever garbage is in memory becomes the third dimension. This is obviously problematic when creating a python array from the dimension object.
2. When a dft array is requested from a component that isn't stored due to symmetry, source specifications, etc, the routine used to return a python array size 0. Python doesn't like this, and produces an array size 1 with whatever garbage is in memory there. Now, the routine returns an array with a size of 1 and with a complex 0. This is especially important for adjoint calculations, which need to store all field components and don't have knowledge about which components meep is actually calculating. The current adjoint toolbox will already broadcast the 0 accordingly.
3. I am now using PyArray_SimpleNewFromData to create and copy the array data.

[smartalecH]

Implemented in #1242.