remove dependency on pytables

[btel]

PyTables is very specialised dependency and it may be hard to install on some systems. It may increase the "entry level" for perspective user. Replacing PyTables with memmapped numpy should not affect performance too much, so pytables can be dropped in future versions.

[btel]

1. spike_sort.extract.filter_proxy implementation should be changed to use memmapped numpy arrays instead of pytables (check performance)

2. The dependency in the IO library should be made optional.

3. Code examples should be modified to use generic binary files (npz, pickle or Bakerlab) instead of hdf5.

4. Documentation and tutorials must be updated accordingly.

[btel]

We might investigate whether PyTables might be replaced by lighter and Py3-compatible h5py library.