h5extract is too slow

Testing on Orca4 jsieren mupage data, it turns out that the new h5extract is slower than the old tohdf5 by a factor of 20. This makes it practically unusable for large scale applications, so some speeding up is required.

Here is the timings of each module using h5extract:

403 cycles drained in 148.415606s (CPU 148.084299s). Memory peak: 447.19 MB
  wall  mean: 0.367872s  medi: 0.335088s  min: 0.173658s  max: 12.586694s  std: 0.609882s
  CPU   mean: 0.367136s  medi: 0.334728s  min: 0.168863s  max: 12.566205s  std: 0.608898s
OfflinePump - process: 0.104s (CPU 0.105s) - finish: 0.001s (CPU 0.001s)
  wall  mean: 0.000259s  medi: 0.000250s  min: 0.000242s  max: 0.000690s  std: 0.000032s
  CPU   mean: 0.000259s  medi: 0.000251s  min: 0.000242s  max: 0.000690s  std: 0.000032s
StatusBar - process: 0.000s (CPU 0.000s) - finish: 0.000s (CPU 0.000s)
  wall  mean: 0.000026s  medi: 0.000025s  min: 0.000025s  max: 0.000030s  std: 0.000002s
  CPU   mean: 0.000041s  medi: 0.000040s  min: 0.000039s  max: 0.000045s  std: 0.000002s
OfflineHeaderTabulator - process: 0.099s (CPU 0.100s) - finish: 0.000s (CPU 0.000s)
  wall  mean: 0.000247s  medi: 0.000236s  min: 0.000229s  max: 0.001621s  std: 0.000072s
  CPU   mean: 0.000247s  medi: 0.000237s  min: 0.000229s  max: 0.001622s  std: 0.000072s
EventInfoTabulator - process: 1.929s (CPU 1.928s) - finish: 0.000s (CPU 0.000s)
  wall  mean: 0.004785s  medi: 0.004584s  min: 0.004533s  max: 0.007842s  std: 0.000543s
  CPU   mean: 0.004785s  medi: 0.004583s  min: 0.004533s  max: 0.007838s  std: 0.000543s
Offline - process: 1.731s (CPU 1.716s) - finish: 0.000s (CPU 0.000s)
  wall  mean: 0.004296s  medi: 0.004004s  min: 0.003961s  max: 0.027509s  std: 0.001283s
  CPU   mean: 0.004257s  medi: 0.004003s  min: 0.003962s  max: 0.012373s  std: 0.000684s
MC - process: 1.987s (CPU 1.986s) - finish: 0.000s (CPU 0.000s)
  wall  mean: 0.004931s  medi: 0.004744s  min: 0.004704s  max: 0.010127s  std: 0.000496s
  CPU   mean: 0.004929s  medi: 0.004743s  min: 0.004705s  max: 0.010110s  std: 0.000495s
MCTracksTabulator - process: 5.193s (CPU 5.191s) - finish: 0.000s (CPU 0.000s)
  wall  mean: 0.012886s  medi: 0.009579s  min: 0.009454s  max: 0.429661s  std: 0.022169s
  CPU   mean: 0.012882s  medi: 0.009576s  min: 0.009452s  max: 0.429575s  std: 0.022164s
RecoTracksTabulator - process: 131.721s (CPU 131.673s) - finish: 0.000s (CPU 0.000s)
  wall  mean: 0.326852s  medi: 0.297362s  min: 0.132525s  max: 12.078222s  std: 0.586519s
  CPU   mean: 0.326731s  medi: 0.297258s  min: 0.132490s  max: 12.073480s  std: 0.586288s
HDF5Sink - process: 5.463s (CPU 5.235s) - finish: 0.077s (CPU 0.070s)
  wall  mean: 0.013556s  medi: 0.013056s  min: 0.009579s  max: 0.030919s  std: 0.002225s
  CPU   mean: 0.012991s  medi: 0.012871s  min: 0.009580s  max: 0.030428s  std: 0.001594s

RecoTracksTabulator takes up by far the longest time, so this is a good place to start.

The double python loop is especially expensive, so I replaced it with vectorizable operations. This leads to a speed up of factor ~6.5 - better, but still not as fast as tohdf5.