Document PyKEEN's performance tweaks

[cthoyt]

Closes #55

This documentation should be a high-level explanation to non-technical users about what is in PyKEEN that makes it better than others. First thing to start with would be copying over the stuff written for the software paper by @lvermue

• Aligned entity/relation ID and embeddings vector position
• Automatic memory optimization
• Sub-batching
• Improvements enabled by separate implementations of score_ht, score_rt, score_hr, score_hrt
• Filtering with index-based masking
• Fast OWA(Not implemented yet)

Any other ideas?

[lvermue]

@mberr I think that this covers most of what we've done so far.
There definitely has to be done some fine-tuning to the formatting and notation.
Anyway, any comments are welcome :)

@cthoyt Your scrutiny is welcome as well :)

[mberr]

I revised the first part about the TriplesFactory / ID-based triple representation. It would be nice to properly link to the properties of TriplesFactory such as entity_label_to_id, etc.

[mberr]

I am not sure whether this is the correct place, but somewhere we should also highlight the risk of manually modifying label-to-ID mappings, and the necessity of keeping the mappings consistent between train/test/validation.

[cthoyt]

@mberr ive written a bit about this in the "bring your own data" tutorial and also made some improvements to it in #54 that haven't yet been merged

[cthoyt]

@lvermue I made edits for clarity and improved the notation used to be a bit more consistent. Maybe you want to look into the algorithm at the bottom of "Filtering with Index-based Masking" to decide if you think this needs more notation, or if it's okay just as text

[cthoyt]

@mali-git @mberr thanks for looking at this, but lets still wait for @lvermue to see if he's happy