Update NOTICE: replace T5Gemma/Flux VAE with Qwen3-VL-Embedding-2B

Remove vae entry (PRXPixelPipeline no longer accepts a vae)

Replace with best-model checkpoint weights (bf16 transformer)

Add PRXPixel 1024 text-to-image pipeline (bf16 transformer, Qwen3-VL text encoder, pixel space)

initial commit