fix: add Apple Silicon (MPS) support (was silently falling back to CPU)

[gaaaork]

Problem

Every device check in nodes.py and encoder_loader.py was torch.cuda.is_available(),
with no MPS branch. On a Mac, that resolves to False, so the pack silently falls back to
CPU for the whole 5B transformer — no error, no warning, just very slow generation.

Confirmed on real Apple Silicon hardware (M-series, 48GB unified memory, PyTorch 2.13):
~1432s/step (24 min) on CPU vs ~4-6s/step once routed to MPS — roughly a 250-400x
difference, same weights, same output.

Fix

Added device_compat.py (new, small, self-contained):

• best_device() — cuda > mps > cpu, replaces the "cuda" if torch.cuda.is_available() else "cpu" pattern used at the one spot that actually decides where the transformer runs.
• empty_cache(device) — MPS/CUDA-aware cache clear, replaces the scattered
  if torch.cuda.is_available(): torch.cuda.empty_cache() calls.

Left untouched on purpose: the CUDA-specific VRAM-budget block-residency planner
(_plan_residency, using torch.cuda.mem_get_info()). MPS's unified memory model doesn't
have a direct equivalent, and the existing "not CUDA -> keep everything resident" fallback
already does the right thing once device correctly resolves to "mps" instead of "cpu" —
no new logic needed there.

Testing

Real end-to-end runs on Apple Silicon, standard Rebels SeFi Loader -> Rebels SeFi Sampler ->
SaveImage graph, weight_dtype=bf16 (fp8_e4m3fn is CUDA-oriented, not exercised here):

• 4-step smoke test: completed cleanly, confirmed Device: mps in the ComfyUI log, no
  unsupported-op errors.
• Real 50-step Base generation (SeFi-5B-Base, 1024x1024, real prompt): 349s total, correct
  image, no errors.

No behavior change on CUDA or CPU-only machines — best_device() returns the exact same
value torch.cuda.is_available() would have on those platforms.

Scope

Deliberately minimal — 3 files, one new. Doesn't touch the folder-structure issue (#1),
unrelated to this fix.