Implement multi-class log-loss

[accenture]

Implement multi-class log-loss

[sumitpai]

Able to achieve MRRs close to table 4 in Rotate paper with multiclass loss. Need further benchmarking before merging.

[rorymcgrath]

In order to make 'generate_corrputions_for_fit' compliant my general understanding is that the following changes need to be made.

1. Update the method signature to take an additional string argument corrupt_side which defaults to 's+o' which can take 3 values, 's' 'o' or 's+o'
2. Add the following logic:

        if 's+o':
            we corrupt subjects and objects. This is the current behaviour
        if 'o':
            we corrupt objects only. Subjects are fixed
        if 's':
             we corrupt subject only. Objects are fixed.

Is this correct?

[rorymcgrath]

Support has been added to generate_corruptions_for_fit assuming the above logic is valid.

[lukostaz]

Has this been resolved and merged to dev?

[sumitpai]

Observation: Performance (MRR) better than nll/pairwise loss over FB15k but does not beat self_adversarial. This was tested with limited parameter combinations. Need extensive hyperparameter search to compare with self_adversarial.