Refactor loading autocorrelation matrix from xml

@manoj.kovalam and I recently tried to understand the function cuda_postcoh_autocorr_from_xml. The style is old, and it was generally hard to follow.

The function loads the autocorrelation matrix from each template, and writes it to GPU.

Changes

I've tried a refactor, mostly some rearranging to group each step more consistently and rename variables a bit more clearly (and hopefully not too verbosely).

A few points of interest:

1. The 'fname' read in is actually a string of the form "<ifoname>:<bank filepath>,<ifoname2>:<bank filepath2>,<etc.>". I've tried calling it ifo_banknames.
2. ifo_banknames is parsed a little differently now. It now checks whether a recognized IFO name is used, but looks a Little more opaque and technical. I think it's in a self contained block so it should be ok.
3. The full process of reading & writing to GPU is done one IFO at a time. The result is a single long function rather than a few clear steps. The only reason to do it that way is to save on CPU memory, reducing the total by a factor of the number of IFOs. As it is, the function uses sizeof(COMPLEX_F) * num_templates * autochisq_len, which is 8 * 1000 * 351 = 2.8MB. It's large enough that I think it's not worth refactoring at least.
4. I added an explicit check that all IFOs have the same number of templates and autochisq length.

Testing

I've tested on an 8000s injection test, and get the exact same results.