We should be able to add an inference-level g2p processing script without mucking around in the model

### Description & Motivation

Here's a situation where:

a) a dataset filelist already contains pre-tokenized phones. First thing is we should better support this.
b) I trained a model using the pre-tokenized phones
c) for inference, we need to add the g2p engine that made these phones, but the engines are defined in the checkpoint

Basically, I think there needs to be some way to hook into the demo or synthesize at inference time, or update the config more easily

Start by tracing the demo and synthesize inference paths described in the issue, along with how checkpoint-defined G2P engines and pre-tokenized phones are currently handled. The work is done when inference can select or provide a G2P engine externally and support pre-tokenized phone filelists without modifying the model.

Written by the indexing model from the issue text.