Support many speakers and improve overall runtime

[ShakedDovrat]

Earlier this year I used your code to create LibriMix datasets for 5, 10, 15 and 20 speakers. In the process I found out that running the scripts for many speakers (10 or more) is very slow or even stuck indefinitely. While debugging I found a bug that caused this and made a few other runtime improvements. The details are listed in the README file and in the code.

I understand that there are a few modifications that need to be made before this PR can be accepted, and I will do so after I get a positive response from a maintainer.

Cheers