Fix training related failures

[JulienVig]

Fixes multiple relatively small bugs happening when training with DISCO.

• Fixes Fix local test failure #612
  One local test was failing with a timeout on my laptop. It happens that the test timeout was too short for my humble machine to finish the test. I simply increased the timeout value.

• Fixes Training on Titanic has a NaN loss #615
  The training Loss was NaN when training the Titanic task. There were two issues:

  1. the tabular preprocessing was empty after a preprocessing (cf. issue for more details). I added a placeholder preprocessing to handle missing values and ensure there is no NaN in the inputs
  2. the default learning rate was too high for the small amount of data and led the weights to diverge. Decreased it from 0.01 to 0.001.
     Before: training accuracy constant at 0.61 and NaN Loss, now the training accuracy increases throughout the epochs (as well as the validation) while the loss decreases monotonically, both from a Node.js script and from the web-client.

• Fixes Train Collaboratively raises error during training #610
  See issue for detailed description of bug and fix. While the client used to send the weights and then request the new weights separately, the server now automatically sends back the new weights after receiving the local contribution. This change saves 1 RTT and importantly, ensures that events happen in the right order, which was previously causing timeouts.
  Now training collaboratively doesn't timeout anymore. The training accuracy now increases monotonically to 70% while it was previously staying constant at 61% with NaN loss.

[tharvik]

small non-blocking comments, merge at will 🚀