Pre-allocate arrays to read multi-module detector data

[takluyver]

This should improve performance by avoiding concatenating large arrays, which involves copying a lot of data in memory. cc @kakhahmed .

Some of the code for this is shared with the virtual CXI writing, as they're now doing a similar task - identifying blocks of source data which can be directly mapped to blocks of the output space.

[takluyver]

This brings the time for getting data through AGIPD1M.get_array() in line with getting data for a single source:

Edit: This is what the timings looked like before:

Note the high CPU time with the AGIPD1M interface, indicating that it was using multiple cores. @kakhahmed did a similar test where it used 1 core and took 40 seconds wall time. I'd guess the difference is in different builds of NumPy or different underlying libraries (like MKL or BLAS). On this branch, the example takes less wall time and less CPU time.

[takluyver]

If you use pulse selection, this selects pulse indices from the file with a numpy array ('fancy indexing'). I suspected that this was inefficient, so I spent a while playing with identifying regular selections and using an h5py MultiBlockSlice (representing an HDF5 hyperslab selection) to process them, in branch xtdf-slice-pulses.

However, it doesn't seem to make a big difference, in testing with some LPD data @daviddoji pointed me to. So I'm leaving it out for now as extra complexity.

[daviddoji]

LGTM!