New descriptors and faster Insight loading

[matteobecchi]

Related Issues: #108
Requested Reviewers: @andrewtarzia

Three main things:

1. Added a new "descriptors" module, which contains miscellaneous, "standard" single-particle descriptors. I put many_body_tica() in it, and also implemented the velocity alignment and the orientational order parameter, commonly used when studying colloidal particles or liquids.
2. Added functions for computing Shannon entropy with KL estimator (useful because aparametric), and negentropy.
3. Changed the way Insights and ClusterInsights are saved/loaded to/from JSON, using an auxiliary.npy file that can be open with mmap_mode, which is way faster fore large datasets.

I think I changed all the related examples and documentation accordingly.

[andrewtarzia]

Mostly minor changes - great work @matteobecchi

[andrewtarzia]

Just added a comment

[matteobecchi]

Addition: a "delay" parameter to the LENS functions, analogously to timeSOAP.

[SimoneMartino98]

all ok for me; just a question / doubt on the convention used for the documentation; maybe it's a silly one, feel free to ignore me.

[matteobecchi]

For future reference: apparently, tSOAP and orientational order parameter are particularly sensitive to numerical errors.

tSOAP because it's a sum of products of hundreds of terms, sometimes very small.

Orientational OP probably because summing complex numbers along the unitary circle, they can be of the order of 1 but their sum very small if they are randomly distributed.