Feature/hdf5 performance optimization

[peter-madigan]

Ok so this is a big PR, but here's the summary:

• Drops python2 support to use a more recent h5py version

• Extends hdf5format to use multiprocessing for the slow conversion of Packet objects to numpy arrays

• Adds a cached int representation for Packet_v2 objects to speed up attribute access

• Introduces a new rawhdf5format for writing raw bytes to disk

• Implements a means of bypassing the packet interpretation and write directly to the rawhdf5format from the PACMAN_IO class

[peter-madigan]

To get a sense of the speed up,

writing 100000 single packet messages (eq. to writing 200000 packets)

In [1]: from larpix.format import rawhdf5format; msgs = [b'header\x00\x00'+b'bodydatadatadata']*100000; rawhdf
   ...: 5format.to_rawfile('test.h5',msgs=msgs)                                                               
In [2]: %timeit rawhdf5format.to_rawfile('test.h5',msgs=msgs)                                                 
746 ms ± 81.2 ms per loop (mean ± std. dev. of 7 runs, 1 loop each)

so ~268000 packets / second

writing 10 length 10000 packet messages (eq. to writing 100001 packets)

In [3]: msgs = [b'header\x00\x00'+b'bodydatadatadata'*10000]*10                                               
In [4]: %timeit rawhdf5format.to_rawfile('test.h5',msgs=msgs)                                                 
57.2 ms ± 4.66 ms per loop (mean ± std. dev. of 7 runs, 10 loops each)

so ~1754000 packets / second

compared to before, writing 100000 packets

In [2]: import larpix; from larpix.format import hdf5format; packets = [larpix.Packet_v2()]*100000; hdf5format
   ...: .to_file('test1.h5',packets)                                                                          
In [3]: %timeit hdf5format.to_file('test1.h5',packets)                                                        
5.43 s ± 178 ms per loop (mean ± std. dev. of 7 runs, 1 loop each)

so ~18000 packets / second

that's a factor of 100