Different upstream and downstream learning rates

[xt0r3-Cambridge]

Implemented support for different upstream learning rates as raised in a feature request.

Use instructions:
Simply add another upstream_lr parameter to the optimiser settings, like this:

optimizer:
  lr: 3.e-4
  upstream_lr: 1.e-5
  [other settings]

The upstream training only happens if the code is run with the --upstream-trainable (or -f) flag.