Add SCAIL-Pose2 / SCAIL-2 workflow support

[rookiestar28]

Summary

This PR adds native SCAIL-Pose2 / SCAIL-2 conditioning support for WanVideo workflows, including replacement and animation dual-mode routing, context-window handling, replacement mask/sample initialization fixes, and a new example workflow.

It also includes a small NLF bbox formatting fix so multi-person bbox candidates are preserved instead of collapsed.

What changed

SCAIL-Pose2 / SCAIL-2 integration

• Added SCAIL-2 loader detection for native pose/mask embedding weights.
• Added SCAIL-2 routing helpers for sampler/model integration.
• Added native SCAIL-2 forward planning for:
  • reference latents and masks
  • additional references
  • pose latents
  • driving masks
  • replacement-mode RoPE/cache planning
  • control strength scaling
  • context-window slicing
• Added WanVideoAddSCAIL2ConditionEmbeds to materialize SCAIL-Pose2 payloads into WanVideo image embeds.
• Routed SCAIL-2 conditioning through WanVideoSampler into the Wan model forward pass.

Replacement mode stability

• Preserved replacement condition structure instead of sanitizing away required pose/reference information.
• Kept raw driving-video condition routing for replacement workflows.
• Added conservative binary latent-mask resizing for SCAIL-Pose2 replacement masks so thin subject masks survive latent downsampling.
• Added spatial and temporal mask growth for replacement masks.
• Added mask-aware samples initialization:
  • preserved/background regions come from encoded driving-video samples
  • replacement subject regions remain random/noised as intended
• Added context-window alignment for samples/noise masks so long-video windows stay frame-aligned.

Dual-mode samples behavior

• Centralized SCAIL-Pose2 samples-disable metadata and helper logic.
• WanVideoEncode.samples can remain wired in dual-mode workflows.
• In replacement mode, samples stay active for replacement/background preservation.
• In animation and other non-replacement modes, the samples payload is marked disabled and ignored by the sampler before video-to-video samples handling.
• Documented this behavior in the README, including the scope of add_noise_to_samples.

Workflow / docs

• Added example_workflows/wanvideo_2_1_14B_SCAIL2_replacement_and_animate_dual_mode_example_01.json.
• Updated README with SCAIL-Pose2 dual-mode samples guidance.

NLF bbox

• Added bbox formatting support that preserves multiple NLF person candidates.
• Added regression coverage for the multi-person bbox output shape.

Tests

Local validation run:

python -m unittest discover -s tests -p "test_*.py"