Fix Phy reading of unit 0 in case of unit exclusion

[alejoe91]

Fixes https://github.com/SpikeInterface/spikeinterface/issues/814

This fixes a bug in read_phy: when excluding noisy, the spike clusters and times were not filtered accordingly, so the spike trains had many additional "0" spikes (because of np.zeros). Now the new spike clusters and times are pre-computed based on remaining unit ids.