Faster unit selection sorting.

[samuelgarcia]

@grahamfindlay @alejoe91

Faster unit selection sorting.

This PR is also a base for discussion to avoid the proposal of get_unit_spiketrains() in #4502
I am not super in favor of adding new internal/external API for the sorting object.

Here this is a simple loop mimic the get_unit_spiketrains()

The speed is not so bad because only a few unit are selected.

phy_folder = Path('/home/samuel.garcia/Documents/data_sorting_graham/')

# no cache
t0 = time.perf_counter()
sorting = si.read_phy(phy_folder)
t1 = time.perf_counter()
print('read phy 342', t1 - t0)
sorting

# no cache
t0 = time.perf_counter()
subsorting = sorting.select_units(unit_ids=sorting.unit_ids[::10])
t1 = time.perf_counter()
print('select 34 units', t1 - t0)

# Trigger cache of spike vector in parent
# Trigger cache of spike vector subsorting
# Trigger lexsort in subsorting but no in parent
t0 = time.perf_counter()
spikes = subsorting.to_spike_vector()
subsorting.to_reordered_spike_vector()
t1 = time.perf_counter()
print('force cache', t1 - t0)

t0 = time.perf_counter()
for unit_id in subsorting.unit_ids:
    st = subsorting.get_unit_spike_train(unit_id=unit_id)
t1 = time.perf_counter()
print('get some spike train', t1 - t0)

give

read phy 342 11.468396085314453
select 30 units 0.030592769384384155
force cache 19.56562849227339
get spike train 0.0018834266811609268

This Pr depend on #4579

[samuelgarcia]

remove this