Support native-layout Krea2 checkpoints (community GGUFs) in from_single_file()

[spartanz51]

What does this PR do?

Builds on #14126 (includes its commits, happy to rebase once it lands).

The converter there covers the repack layout, but the community GGUF exports of Krea 2 (vantagewithai/Krea-2-Turbo-GGUF, realrebelai/KREA-2_GGUFs, ~145k downloads combined) keep the native key names from the krea-ai/krea-2 inference code: prenorm/postnorm.scale, attn.qknorm.*, flat mod.lin tables, first/tmlp/tproj/txtmlp/last.*. Those fall through remap_key untouched, so loading ends with ~180 unassigned params and crashes on meta tensors.

This extends the converter:

• norm scales → norm1/norm2 and attn.norm_q/norm_k
• flat mod.lin → per-block scale_shift_table, last.modulation.lin → final_layer.scale_shift_table
• first/tmlp/tproj/txtmlp/last.* → img_in/time_embed/time_mod_proj/txt_in/final_layer
• drops last.up/down.weight (not part of the current LastLayer; the official diffusers weights don't ship them either — 430 keys vs 432 in native exports)
• reshapes the squeezed projector.weight found in some exports

Tested with GGUFQuantizationConfig on Q3_K_M / Q4_K_M / Q8_0 from both repos: all load with zero missing/unexpected keys, and end-to-end Krea2Pipeline generation (8 steps, 768x768, mps) produces correct images.

Fixes the GGUF side of #14122.

Who can review?

cc @buffett0323 @DN6

[github-actions]

Hi @spartanz51, thanks for the PR! It does not appear to link an issue it fixes. If this PR addresses an existing issue, please add a closing keyword (e.g. Fixes #1234) to the PR description so the issue is linked. See the contribution guide for more details. If this PR intentionally does not fix a tracked issue, a maintainer can add the no-issue-needed label to silence this reminder.