Add new percentile aggregator and tests for comparing it with Byzantine

[mina5rovic]

Implementation of new percentile aggregator based on the previously implemented Byzantine (in previous version of disco). This one work faster and in some cases better than the Byzantine. Also implemented tests for comparing those two.

Changing Byzantine aggregator based on errors found in implementation while comparing them.

[JulienVig]

The comparison is very useful, thanks!

I have a bunch of questions, mostly due to my superficial knowledge of the algorithms. Overall, we're often getting exact values like 1.0000 which seems a bit fishy. Also, the new aggregator seems worse than the old both time- and robustness-wise. I believe it should at least be more robust so there are maybe some bugs that slipped into the implementation?

Finally, could you add a memory test case to ensure that the implementation is currently doing garbage disposal of unused TF tensors? You can do so by running multiple rounds/aggregations and keeping track of the number of tensors allocated in memory (tf.memory()), which ideally should stay constant throughout the training. You can read more on memory management here.