Better defaults for chunks parameters

[jonbrenas]

Addresses #629.

This is just a prototype to check that it is what @alimanfoo had in mind. I have not tested that it actually improves performances yet.

If that's the correct idea, I will extend that to other similar functions.

[leehart]

@alimanfoo

[leehart]

I'll try re-running the integration and notebook tests here, to see if it automatically picks up the fix for gcsfs.

Ref: issue #782

[leehart]

By the way, did we ever find out whether this was what was in mind? I wonder if there is a way to confirm, e.g. if performance is improved and nothing breaks and it seems to make sense to others, etc.

[jonbrenas]

I am not sure that was exactly what @alimanfoo had in mind. I think there is a piece of chunk optimisation in the PCA part of the cohort analysis which is what he meant to do but I have not checked that it does the same thing as this. I'll move this PR back to draft to force myself to remember to check it.