Fix AuraFlow VAE dtype mismatch on pipeline reuse (#14184)
upcast_vae() upcasts the whole VAE to float32 in place, so on a second
__call__ needs_upcasting is False and the latents cast guarded by that
branch is skipped, feeding fp16 latents to the fp32 VAE. Cast the latents
to the VAE dtype inline at the decode call so it always runs, matching the
fix applied to pixart_sigma in #8391.
Co-authored-by: IvenHsu01 <228416682+IvenHsu01@users.noreply.github.com>
Co-authored-by: dg845 <58458699+dg845@users.noreply.github.com>
