Singleton DFT fix

[smartalecH]

Fixes #1269

There were actually two bugs here. First, a collapsed array size counter (reduced_dims) was uninitialized. If the array to be collapsed only consisted of a single element, it wouldn't update to the correct value. Consequently, the routine would try to create an array the size of whatever garbage was in memory (which would occasionally cause a segfault).

Next, the actual singleton DFT calculation was never copied to the garbage array (assuming no segfault occurred).

Note there are still probably some other bugs in the collapse_array() routine (array_slice.cpp). If we see more odd behavior with DFT fields (e.g. when symmetries are applied, weird field components, etc.) we should first look here. Then we should look in _get_dft_array() inside of meep.i.

[smartalecH]

As discussed I check for singleton cases earlier and return.

However, I noticed there is another bug that still requires the initialization of reduced_dims.

Taking the simple example in #1269, if the user specifies a line with size (1,22,1) (i.e. two dimensions need to be collapsed), then the reduced stride is set to garbage:

meep/src/array_slice.cpp

Lines 678 to 680 in 5ad4b84

	else if (reduced_stride[1] != 0)
	reduced_stride[1] = reduced_dims[1];
	// else: two degenerate dimensions->reduced array is 1-diml, no strides needed

The comment in the code seems to assume that the reduced stride would be set to zero by default.

[stevengj]

The later code should be:

if (reduced_rank == 2) {
    if (reduced_stride[0] != 0)
      reduced_stride[0] = reduced_dims[1];
    else if (reduced_stride[1] != 0)
      reduced_stride[1] = reduced_dims[1];
}

Basically, it should never access elements of reduced_dims ≥reduced_rank

Even simpler:

if (reduced_rank == 2)
    reduced_stride[reduced_stride[0] != 0 ? 0 : 1] = reduced_dims[1];

[smartalecH]

Fixed -- should be ready for merging after tests pass.