Please update the code to use latest Kokoro models (non V0 models)

[finnvyrn]

Currently ./docker/rocm/docker-compose.yml downloads models like bf_v0isabella, which sounds significantly worse than bf_isabella, found in projects like https://github.com/nazdridoy/kokoro-tts

[RBEmerson970]

I concur - the V0 voices are pretty much useless.

[remsky]

Fixed on master, should be in the next baked release re: #480

Thanks for the catch. The v0* voices are the old Kokoro v0.19 packs kept around as legacy, and their embeddings sound degraded on the v1.0 model. If I remember right there was some adaptation needed to get them working at all, but they were added on request, for continuity despite the quality issues.

The v1.0 bf_isabella was missed in the bulk swap of voice packs when we upgraded the base model/library for v1.0. Think the rest should all be fine and accounted for already, but let me know if there's something missed. Also noticed that the fallback mappings for the OpenAI endpoints were still pointing at the old ones in some cases and have that fixed now.

Changes:

• added the v1.0 bf_isabella
• repointed nova -> bf_isabella, plus alloy/ash/coral/echo off the legacy v0 voices onto the v1.0 versions in the openai default mapping.
• the v0* voices will still stay available, but only when explicitly called.