The HWInference process' footprint was more or less entirely unaccounted for. The size comes from the loader, through a new parakeet_capi_weights_bytes(), and is kept in one process-wide number: only one session runs at a time. We'll need to do something similar for llama.cpp, but it's clearer labeled as such, tying the memory usage to the feature. Differential Revision: https://phabricator.services.mozilla.com/D324040