Krea 2 regional prompting

[JPPhoto]

Summary

Adds regional prompting support for Krea-2 using Flux-style restricted attention.

This PR:

• Allows masked Krea-2 conditionings and conditioning collections.
• Adds positive regional prompting to the canvas.
• Applies enabled conditioning enhancers independently to global and regional prompts.
• Accounts for retained attention masks and mask-construction scratch memory.
• Adds CUDA-gated coverage for dense masks with memory-efficient SDPA.
• Documents dense-mask memory costs and math-attention fallback behavior.

Related Issues / Discussions

Builds on #9406.

QA Instructions

For manual canvas QA:

1. Select a Krea-2 model.
2. Add a positive regional prompt.
3. Enable Conditioning Rebalance and Seed Variance.
4. Generate and confirm both global and regional conditionings pass through separate enhancer chains.

Merge Plan

Checklist

• The PR has a short but descriptive title, suitable for a changelog
• Tests added / updated (if applicable)
• ❗Changes to a redux slice have a corresponding migration
• Documentation added / updated (if applicable)
• Updated What's New copy (if doing a release after this PR)

[Pfannkuchensack]

Findings

• Medium: invokeai/app/invocations/krea2_denoise.py:521 (_regional_attention_mask_bytes) under-reserves working memory because it ignores the float attention bias PyTorch materializes from the boolean mask on every SDPA call.
  Chain: invokeai/backend/krea2/regional_prompting.py:90 allocates the mask as torch.bool, and attention_mask_numel (invokeai/backend/krea2/regional_prompting.py:51-56) reports total_seq_len**2, i.e. exactly the bool byte count. That value plus attention_mask_build_scratch_numel is the whole of the regional contribution added to _estimate_working_memory at invokeai/app/invocations/krea2_denoise.py:566. But the mask is handed to F.scaled_dot_product_attention(..., attn_mask=attention_mask) at invokeai/backend/krea2/attention.py:87; PyTorch converts a boolean attn_mask into an additive bias in the query dtype before dispatch, so each masked attention transiently allocates a further total_seq_len**2 * 2 bytes (bf16) that nothing reserves.
  Evidence (measured on RTX 4090, torch 2.7.1+cu128, 48 heads / head_dim 128 / bf16, seq 4608): unmasked SDPA peak 54 MiB, masked peak 94 MiB, bool mask itself 20 MiB. The 40 MiB delta is exactly 4608**2 * 2. At 2560x1440 with ~400 text tokens (total_seq_len ~14800) that is ~440 MB unaccounted, on top of the ~219 MB the estimate does count per mask.
  Trigger: any regional generation at high resolution on a card where the cache places the transformer using the returned hint. The estimate carries several GB of deliberate slack (invokeai/app/invocations/krea2_denoise.py:549-554), which is why this is Medium and not High, but the PR's stated contribution is exact mask accounting and this is the largest mask-derived allocation.
  To expose this issue, add a test that asserts the regional byte estimate scales with the inference dtype width (that it includes total_seq_len**2 * dtype_bytes for the live mask in addition to the bool storage), or a CUDA-gated test that compares torch.cuda.max_memory_allocated() for one masked Krea-2 attention against _regional_attention_mask_bytes.

• Low: invokeai/backend/krea2/regional_prompting.py:92-104 silently drops the global/background prompt when the region masks cover every grid cell, and no validator warns about it.
  Chain: background_mask starts all-True and is ANDed with ~region for each masked conditioning (line 95). If the union of regions covers all image_seq_len cells it becomes all-False. For an unmasked (global) conditioning the code then writes that all-False vector into both the text-to-image block (line 103) and the image-to-text block (line 104), so no image token attends to the global prompt and the global prompt attends to no image token; only its own diagonal text block survives. Confirmed by direct probe: with one unmasked conditioning plus an all-ones region mask, mask[global_range, image_block].any() and mask[image_block, global_range].any() are both False.
  Trigger: on the canvas, invokeai/frontend/web/src/features/nodes/util/graph/generation/buildKrea2Graph.ts always wires the global pos_prompt encoder into pos_cond_collect, so a user who draws a regional guidance layer over the whole canvas (or, at low resolution, a layer that covers every 16 px grid cell) gets their main prompt entirely ignored with no feedback. getRegionalGuidanceWarnings in invokeai/frontend/web/src/features/controlLayers/store/validators.ts:104-116 does not flag it.
  This is FLUX parity (invokeai/backend/flux/extensions/regional_prompting_extension.py:127-208 behaves identically), so it is not a regression - but Krea-2 ships it alongside a docs claim that an unmasked conditioning "applies to the background not covered by any regional mask", which is vacuous in this case.
  To expose this issue, add a test that builds an extension from one unmasked conditioning plus an all-ones region mask and asserts the intended fallback (background coverage restored, or an explicit error), rather than the current silent no-op.

[JPPhoto]

@Pfannkuchensack I think everything you raised is addressed.

[Pfannkuchensack]

Works.