⬇️🎖️ Low rank embeddings

[mberr]

This PR adds a representation module, which factorizes the Embedding matrix into a matrix of basis representations, and trainable linear weights.

It is an alternative way to reduce the number of trainable parameters, while keeping the embedding dimension.

• It is essentially the same as
  

  pykeen/src/pykeen/nn/message_passing.py

  Lines 141 to 150 in c8f6af6

  	class BasesDecomposition(Decomposition):
  	r"""Represent relation-weights as a linear combination of base transformation matrices.
  	The basis decomposition represents the relation-specific transformation matrices
  	as a weighted combination of base matrices, $\{\mathbf{B}_i^l\}_{i=1}^{B}$, i.e.,
  	.. math::
  	\mathbf{W}_r^l = \sum \limits_{b=1}^B \alpha_{rb} \mathbf{B}^l_i
  	"""

so there may be some potential to reduce duplication. -> 78d9b62

[mberr]

I had to move this one to avoid cyclic imports

[cthoyt]

🚨🚨 IT'S THE DOCUMENTATION POLICE, OPEN UP 🚨🚨

besides the notes on improving the docs, I think this all looks solid. the moves for the circular imports make sense though they make me wonder if we should adopt pytorch-like way of just importing everything in pykeen.nn

[cthoyt]

@PyKEEN-bot test

[mberr]

@PyKEEN-bot test