Update fairseq TTS pipeline

[kahne]

No description provided.

[osanseviero]

Thank you very much for this PR!

This seems to have a couple of issues as seen in the Docker tests during installation of requirement (pkuseg). I was not able to reproduce unfortunately, maybe @Narsil has some ideas otherwise I can dive a bit more into this.

[osanseviero]

Looks good from my side, thanks @Narsil and @kahne!

[Narsil]

It seems the issue lies in the pkuseg that requires numpy to already be installed as it uses it in its installation script.

I took the liberty of modifying
self.generator = self.task.build_generator(self.model, cfg) into self.generator = self.task.build_generator(model, cfg) (first is the raw model, second is a model list)

Also adding another layer of protection against bogus data ("\x00\x00" as an input gets transformed into nil tensor which triggers and error).