Add macOS (Apple Silicon) support for the interactive demo

[zxytim]

Summary

Enables the real-time interactive demo to run on macOS (Apple Silicon) using the MPS backend — no CUDA/TensorRT required. Closes #1.

Changes

Device support

• Motion model runs on MPS (auto-selected; ARDY_DEVICE override); text encoder runs on CPU. Demo launched with --no-compile (TensorRT unavailable on macOS).
• Cast float64 skeleton buffers to float32 (float64 is unsupported on MPS).

MPS thread-safety (fixes hard crashes)

• MPS is not thread-safe: the playback thread (per-frame skinning), the replan thread (generation), and the server thread (model loads, prompt updates) previously issued MPS work concurrently, crashing the Metal driver (command-buffer assertion, or SIGSEGV after extended sessions).
• Model (re)loads and generation steps now run under a shared GPU lock.
• All session/visualization state (motion tensors, joints, contacts, reference motion, skin bind data, mesh rest poses) is kept on CPU, so the per-frame playback path never touches the accelerator.
• Skinning (lbs) follows the pose tensor's device.

Text encoder without gated Llama access

• Load LLM2Vec from local adapter repos (TEXT_ENCODERS_DIR) with the base model repointed at the ungated, byte-identical mirror NousResearch/Meta-Llama-3-8B-Instruct (the default meta-llama base is gated).
• Resolve local PEFT-adapter repos explicitly in the vendored llm2vec.py: transformers>=5 no longer follows adapter_config.base_model_name_or_path when a config.json sits beside it.

Usability

• run_demo_mac.sh: one-command launch; bootstraps text_encoders/ on first run.
• README: macOS setup + launch section.

Testing (macOS 15, M-series, 128 GB)

• Core/Horizon-40 model: ~0.3 s per 2 s generation window on MPS (several times real-time); playback holds native 20 fps with auto-replan.
• Verified stable across: continuous playback, live prompt edits, model reload mid-playback, Core↔G1 skeleton switch (character recreation), target-velocity steering, waypoint placement, and multi-client connect/reconnect.
• Previously reproducible Metal crashes on model reload / long sessions no longer occur.

Notes

• [trt] extra is CUDA-only; macOS installs use pip install -e ".[demo]".
• No behavior change on CUDA: the GPU lock is harmless (re-entrant) and CPU-state conversions are no-ops where tensors were already on CPU.