Add VoxCPM engine and LoRA training

[diodiogod]

Summary

• add the official VoxCPM engine with VoxCPM2, VoxCPM1.5, and legacy VoxCPM-0.5B support
• integrate VoxCPM2 voice cloning, controllable cloning, reference-free Voice Design, TTS Text, and TTS SRT
• add official-format VoxCPM2 LoRA dataset preparation, configuration, training, checkpoint resume, warm-start, discovery, and inference loading
• make LoRA adapter identity part of model and audio cache signatures
• automatically select the CUDA device with the most free VRAM for training, with an explicit device-index override
• update generated engine documentation and model layouts

Why

This addresses the VoxCPM support requested in issues #99 and #198 while using OpenBMB's official implementation as the runtime and training reference. VoxCPM2 is the recommended model; legacy generations remain available for inference, while integrated training is deliberately limited to VoxCPM2 until legacy compatibility is tested.

Validation

• Python compilation passed for all changed modules
• git diff --check passed
• ComfyUI restarted successfully and registered the new engine/training inputs
• a real one-step VoxCPM2 LoRA training smoke run completed on an RTX 4090 in about 38.6 seconds
• the exported lora_weights.safetensors and lora_config.json adapter loaded through the normal VoxCPM engine and generated audio successfully
• safe checkpoint save/load and optimizer/scheduler resume contract passed

Testing status

This is intentionally a draft because broader manual workflow and regression testing still needs to happen when the maintainer is back at the test machine. No version bump is included.