Make running NMT jobs locally more straightforward

[Enkidu93]

A simple way to do this would be to:

• Make a flag that can be passed through to the specify the device (so you can use the CPU).
• Add a flag and alternative translation pipeline that just echos back the source segments.
• Use the NLLB tiny random model.
• Specify a configuration using these flags/this model in the launch.json.