feat: expose img2img CLI flags for Flux.2 Klein

[terribilissimo]

Add --image-path and --image-strength to mflux-generate-flux2, following the same pattern used by flux_generate.py (Flux.1).

Note: This is independent of #353 (Z-Image enhancements), and can be reviewed/merged separately.

Three additions:

1. parser.add_image_to_image_arguments(required=False)
2. reference_image_path=args.image_path in DimensionResolver.resolve()
3. image_path + image_strength passed to model.generate_image()

Summary

The Flux2Klein Python class already implements full img2img support via
_prepare_img2img_latents(), accepting image_path and image_strength
parameters in generate_image(). However, the mflux-generate-flux2 CLI
does not register --image-path / --image-strength, so users cannot
access img2img from the command line.

This PR adds 3 lines to flux2_generate.py to expose these flags,
following the exact same pattern used by flux_generate.py (Flux.1):

1. parser.add_image_to_image_arguments(required=False)
2. reference_image_path=args.image_path in DimensionResolver.resolve()
3. image_path=args.image_path + image_strength=args.image_strength in generate_image()

When --image-path is not provided, behavior is identical to upstream.

Why this is different from mflux-generate-flux2-edit

Flux.2 Klein already has an image-conditioned editing CLI
(mflux-generate-flux2-edit with --image-paths), but that is a
fundamentally different operation from img2img:

	img2img (--image-path)	Image-conditioned editing (--image-paths)
Purpose	Transform a single input image while preserving its overall structure and composition	Use one or more reference images as semantic context for generation
How it works	Encodes the input into latent space, adds noise controlled by --image-strength, then denoises — the result follows the spatial layout of the original	Reference images are injected as conditioning tokens alongside the text prompt — the model draws on their content freely
Strength control	--image-strength (0.0 = ignore input, 1.0 = pure noise) controls how much of the original is preserved	No strength parameter — the prompt determines what changes
Number of inputs	Exactly 1 image	1 or more reference images
Typical use cases	Style transfer, color grading, detail refinement, artistic re-interpretation of a photo	"Put these sunglasses on this person", combining elements from multiple images, subject-driven editing
CLI command	mflux-generate-flux2 (this PR)	mflux-generate-flux2-edit (already exists)

Every other generation command in MFLUX that supports img2img
(Flux.1, FIBO, Qwen Image, Z-Image) already exposes --image-path /
--image-strength in its CLI. Flux.2 Klein is the only model where the
underlying class supports it but the CLI does not — this PR closes that gap.

Testing

Tested on Apple Silicon with flux2-klein-4b (distilled) and
flux2-klein-base-9b (undistilled). Verified:

• txt2img (no new flags) produces identical output to upstream
• img2img with --image-path + --image-strength works correctly
• DimensionResolver auto-sizes from the reference image when --width/--height are omitted

Checklist

• Single file changed (flux2_generate.py)
• No new dependencies
• Backward-compatible (no flags = identical behavior)
• Follows existing pattern from flux_generate.py

[filipstrand]

Should have been there from the start :). Will make a 0.16.5 release with this included.

[terribilissimo]

Thanks!