Load VAE (float32)

Loads a VAE in float32 from the start, as SDR to HDR's float32 encode and decode need. Core's VAE Loader builds it in bf16, and LTXVVAEForceFloat32 then casts it in place: under ComfyUI's dynamic VRAM, its default without --highvram, the buffers stay sized for bf16 and the HDR pass failed on a 32 GB card with « Buffer too small » (2 Nov 2026).