3️⃣🦌 Add TripleRE model

[mberr]

Reference Implementation: https://github.com/LongYu-360/TripleRE-Add-NodePiece/blob/master/TripleRE%2BNodepiece/ogb_wikikg2/model.py

Note: this PR also renames NodePiece's token_representation_kwargs to token_representations_kwargs for consistency.

Tasks:

• support choosing "mlp" aggregation for NodePieceRepresentation (moved from NodePiece model)
• rename GalkinPickleLoader to GalkinPicklePrecomputedTokenizerLoader for consistency
• rename NodePiece's token_representation_kwargs to token_representations_kwargs for consistency
• pass evaluation_factory to sampled evaluator, if this evaluator is used
• allow selection of loader for PrecomputedPoolTokenizer
• allow loading precomputed negative samples for sampled evaluator
• since OGB uses a fixed set of negative samples for (sampled) evaluation, there need to be different evaluator instances for validation / testing, which is currently not supported by the pipeline (closes Build a pipeline for evaluating with OGB #366)

Dependencies:

• ✨🦌 Add TripleRE interaction #712
• ☮️⚓ NodePiece with anchors #755
• NodePiece with joint relation and anchor tokenization #766 / 🍸✌️ Mixed tokenization for NodePiece #770
• MixtureAnchorSelection contains duplicates #772
• 🥒🏴‍☠️ Precomputed Tokenization for NodePiece #822

Fixes #711

[migalkin]

Complete reproduction of their code would require joint Anchor + Relation tokenization 😉

[cthoyt]

@migalkin BAHHH

[migalkin]

I'd be in favor of merging this model without that special OGB evaluator