ACE-Step 1.5 vendored source

- Source: https://github.com/ace-step/ACE-Step-1.5
- Commit: `14c0211d5a0653b0f63e27686f4c3f151b4d8629`
- Imported paths: selected inference files under `acestep/` and the bundled
  `acestep/third_parts/nano-vllm/nanovllm/` runtime
- License: MIT; see `LICENSE`

Only code required by Xinference music inference is retained. Upstream API
servers, CLI/Gradio UI, dataset tools, training implementations, tests,
examples, development metadata, unused DiT variants, and nano-vllm assets are
excluded. Xinference supports only the bundled `acestep-v15-turbo` DiT here.

Local integration patch: remove `LoraManagerMixin` and
`TrainingPresetMixin` from `AceStepHandler`, because Xinference exposes neither
training nor upstream LoRA-management endpoints. General generation, LM,
PyTorch, MLX, model-code synchronization, and audio encoding remain intact.

The retained nano-vllm runtime keeps its separate MIT license at
`acestep/third_parts/nano-vllm/LICENSE`.

Runtime dependencies such as PyTorch, Transformers, Diffusers, and TorchAO are
not vendored. Xinference installs them through the ACE-Step model virtual
environment.
