feat: Apple Silicon (MPS) local generation support

[Sergio Gil Jiménez (jimeneztion)]

Summary

Enables local AI video generation on Apple Silicon Macs. Previously, all macOS
users were forced to use the LTX cloud API regardless of hardware. With this
change, users with ≥15 GB of unified memory can run full local inference on
their GPU via Metal Performance Shaders (MPS).

The work has three layers:

1. Low-level MPS compatibility patches (backend/services/patches/)

The upstream ltx_pipelines library is CUDA-first and makes several API calls
that don't exist on MPS. We monkey-patch at server startup to fix them without
forking upstream:

• mps_gpu_model_fix.py — replaces CUDA stream-based async layer streaming
  with a synchronous MPS-aware wrapper. On CUDA, layers are streamed to GPU
  using async CUDA streams for overlap; on MPS we do it synchronously since
  Metal has no equivalent primitive.
• mps_layer_streaming_fix.py — skips pin_memory() when moving layers to
  MPS. Pinned host memory is CUDA-only; calling it on MPS silently corrupts
  tensors.
• mps_vocoder_fix.py — fixes a float32 dtype mismatch in the vocoder. MPS
  autocast doesn't support float32, so we temporarily cast model weights
  instead of relying on autocast.
• safetensors_loader_fix.py — sets non_blocking=False and copy=False
  when moving memory-mapped tensors to MPS. Async transfers on mmap buffers
  can segfault on Metal.
• ltx_text_encoder.py — inlines device-aware memory cleanup to avoid
  unconditional torch.cuda.synchronize() calls.

2. Local generation policy (backend/runtime_config/runtime_policy.py)

• Before: Darwin → always API-only.
• After: Darwin with ≥15 GB unified memory → local generation allowed.

All four pipelines (fast, a2v, ic_lora, retake) set
streaming_prefetch_count=1 on MPS (synchronous) instead of 2 (async with
CUDA streams). This prevents OOM when streaming the transformer without CUDA
stream overlap.

3. Warmup skip on MPS (health_handler.py, pipelines_handler.py)

A full inference warmup at startup takes several minutes on a cold Metal device
and blocks all generation requests. On MPS we skip warmup — the pipeline loads
and the first real generation acts as warmup. For CUDA, the warmup state machine
is unchanged but now correctly manages WARMING → WARM transitions so requests
that arrive during warmup wait correctly instead of failing.

How to test

1. Apple Silicon Mac (M1/M2/M3/M4) with ≥15 GB unified memory
2. pnpm dev
3. Settings → disable "Use LTX API"
4. Generate a video — should run locally on MPS, no cloud API call
5. Logs should show [t2v] Generation started (model=fast, ...) with no CUDA errors
6. pnpm typecheck && pnpm backend:test