Freezing of CATS_SCORE and SCORE during training with GPU -0 configuration and fill-config CPU settings

[dmesquita]

Hi everyone!

This is the context:
When using init fill-config without specifying the learn_rate, it fills in the training settings for CPU (I still need to look into this further, but from what I've seen in the code, I believe this is the case). When training a pipeline with the textcat component using --gpu-id -0 and this configuration, no error is generated; however, the CATS_SCORE and SCORE remain frozen at each eval_step.

Initially, I thought about creating a pull request with a way to specify CPU or GPU in the init-config command, but it would involve many changes. Then I thought it might be better to check this in the debug config and raise the error, but at the validation stage, we still don't know if the user is trying to train using --gpu-id -0 and a CPU configuration.

What do you recommend I propose to solve this problem? Actually, before that: is it worth solving this, or is it a rare use case? Thank you!

============================== Info about spaCy ==============================

spaCy version    3.7.4 
Platform         Linux-5.15.0-101-generic-x86_64-with-glibc2.35
Python version   3.10.12                       
Pipelines        en_core_web_md (3.7.1), en_core_web_trf (3.7.3), en_core_web_sm (3.7.1)