Untiled LTX VAE decode when it fits, drop per-tile gc

[TroyHernandez]

Summary

The video decode phase measured 12.3 s in-render. Profiling it in isolation (VAE only, synthetic latents at the 768x512x49 shape): the tiled decode was 4.35 s, of which 1.89 s was the explicit per-tile gc(), and a single untiled forward is 2.67 s. Tiling exists to bound activation VRAM, not for correctness — and at decode time the transformer is phase-offloaded, so that VRAM is actually free.

• decode() now runs one full-latent forward when the estimated activation cost fits the card. Cost is linear in output pixel-frames: ~1 GB + ~350 B/pxf, calibrated across 2.5M–19.3M pixel-frames (three sizes, linear to within noise); the rule uses 360 B/pxf plus 15% headroom against total-minus-allocated. options(diffuseR.vae_untiled = TRUE/FALSE) forces either path; "auto" applies only on CUDA and CPU keeps the tiled behavior. 1024x576x121 (~26 GB estimate) correctly stays tiled on a 16 GB card.
• The per-tile explicit gc() in both tiled paths is gone. It guarded against an allocator-callback storm that no longer occurs with the gc gates live (Activate the LTX allocator gates and pinned staging #37): measured 3.8 s vs 4.2 s per decode, storm-free, VRAM peak unchanged.

Measurements

768x512x49, NF4, pinned staging, seed 42, RTX 5060 Ti:

	before (post-#37)	this PR
video decode phase	11.4–13.2 s	1.7–1.8 s
warm render	54.0–54.3 s	44.2–44.5 s

The week's cumulative arc for this render: 89–96 s → 73 s (byte-LUT #35) → 54 s (gates + pinning #37) → 44 s.

Validation

• Full local suite: 892 results, all green, including 3 new dispatch tests (forced-untiled output identical to the untiled reference, forced-tiled identical to the tiled path, auto-on-CPU stays tiled).
• Calibration note: a first calibration attempt read a flat 13.0 GB peak at every size — that was the loader's pool pre-warm polluting the cumulative peak stat, dodged by loading on CPU and moving the VAE manually.