[TroyHernandez]

Adds img2img_flux2(prompt, image = ): image + prompt editing with FLUX.2 klein-4B, using the model's native reference-image path (same weights, no new download). One reference image for now.

What's new

• flux2_vae_encoder() / load_flux2_vae_encoder(): the AutoencoderKLFlux2 encode path (encoder body, quant_conv, BN stats), reusing the decoder's resnet and mid-block modules. Returns the posterior mean (sample_mode = "argmax").
• flux2_encode_image(): encode, patchify 2x2, BN-normalize, pack to tokens.
• flux2_prepare_image_ids(): reference position ids with T = 10 per image (list-shaped for a later multi-image change).
• flux2_preprocess_image(): diffusers' preprocessing (min side 64, aspect <= 8, Lanczos area cap at 1024^2, floor to /16, center crop, [-1, 1]). The Lanczos resize is a port of Pillow's coefficient algorithm as a separable matmul.
• flux2_load_pipeline(encoder = TRUE) loads the encoder; the default leaves txt2img memory unchanged.
• txt2img_flux2() and img2img_flux2() share one internal generator. The reference tokens are appended to every transformer forward and the prediction is sliced back to the target tokens; mu uses the target length only.

Validation (real weights, fp32 CPU, vs diffusers)

• Encoder, 512x512: relative error ~2e-6 at every block.
• Preprocessing: exact at 1024x1024; the 1920x1080 resize path is off by at most 1.1/255 (mean 0.25/255), from PIL's 8-bit rounding.
• End-to-end 256x256 edit, seed 7, shared prompt embeds: PSNR 123 dB.
• 27 new tinytests, including an encoder parity fixture (tiny random VAE, generated by tools/gen_fixtures_flux2_vae_encoder.py). Full suite: 1487 passed, 0 failed.

Not done

• GPU memory/time at 1024x1024 with a reference: not measured (the card was held by the fleet worker). One reference doubles the image tokens, so expect a peak above txt2img's 12.5 GB.
• Fleet support (resident_load, gpu.ctl entry, xtx.api) is a follow-up PR.

Version bump to 0.2.2.15 is in its own commit.