LoRA fallback follow-ups from #32 (TAE on Metal, layer-split probe, test gaps) #33

Non-blocking follow-ups from the re-audit of #32 (CLEAR at 6c4d699; evidence llamadart-evidence/sd-native-pr-32/audit-r2/).

• TAE LoRA silently not applied on Metal. CPU auto merges a tae. LoRA; Metal falls back to at_runtime, which upstream never applies to the TAE. Previously it aborted. Warn (or refuse) when a LoRA has tensors no module consumes. Reachable only via tae.-prefixed files or SDXS VAE LoRAs.
• Layer split probe: the fallback probe uses runtime_backend(), the first device of a & split (ggml_extend_backend.cpp:934); CPU&MTL0 isn't probed on Metal. Use runtime_backends(module).
• Unprobed modules: CLIP_VISION, CONTROL_NET, PHOTOMAKER, AUDIO_ENCODER.
• Tests: BLAS "keeps merging" untested (host-buffer shortcut mutant survives); stale docstring tests/test_fp8_runtime.py:15 (runners list MTL0 but can't generate); CI never exercises the Metal fallback (local only).
• Warning wording: "F16 weights on MTL0" appears for F32 models (accurate: conv weights load as F16) — consider naming the cause.
• Behavior note: linear-only LoRAs on F32/BF16 bases on Metal now apply at runtime instead of merging (same image, per-step cost).