Add spike-train based lazy SortingGenerator

[h-mayorquin]

While discussing in https://github.com/SpikeInterface/spikeinterface/pull/2175 and https://github.com/SpikeInterface/spikeinterface/pull/2209 and the other PR that is now I realized that we don't have a way of tesing lazy recorders performance.

Currently, generate_sorting alredy pre-loads the data as a numpy array which means that we can't test how performancer or other IO operations would behave in different scenarios.

This PR aims to fix that by introduces a new sorter that generates spikes on the fly in a reproducible way. Plus, it behaves consistency across calls and by slicing across spikes.

The max memory it should use is the complete size of its longest spike train.