Implement cross-batch memory for losses

[monatis]

• Paper: https://arxiv.org/pdf/1912.06798.pdf
• Reference for implementation: https://github.com/msight-tech/research-xbm/

How it works

• XBM relies on the observation that the drift of embeddings is slow during training, i.e., embeddings for the same object is changing in a very slow pace.
• This lets us add embeddings and targets in a ring buffer of a certain size.
• After a certain number of iterations, start using the buffer. Now the final loss is the weighted sum of the actual mini-batch loss and the ring buffer loss.

Suggested implementation

1. Introduce an XBMConfig class to hold the configuration values such as buffer_size, start_iteration, xbm_weight.
2. Add a configure_xbm() hook in TrainableModel and return None by default.
3. İf it returns an XBMConfig instance instead, create a XBMBuffer instance in the TrainableModel constructor.
4. Implement the XBM logic in _common_step if stage is training.

Notes

1. We cannot re-use the existing Accumulator classes because they are not ring buffers.
2. I don't think we need a mixin because addition to TrainableModel will be only a few lines of code, and we need to update _common_step anyway.

[monatis]

Suggested implementation is in the issue. WDYT? @generall and @joein

[monatis]

Completed in #175