Fix Lens conditioning capability reporting

Merge remote-tracking branch 'origin/master' into paralol/main-gpu-pi…

…peline-cuda133

Add Lens and Bonsai experimental fork work

Adds Lens/Qwen GPT-OSS conditioning, staged transformer/runtime APIs, smoke tools, and model capability reporting.

Keeps Bonsai GemLite INT1 CUDA spike disabled by default, preserves its ggml custom-op delta as a patch artifact, and avoids adding the CUTLASS submodule as a required dependency.

Build verified with Release CUDA shared library and sd-latent-smoke after resetting ggml clean.

Emit TAE previews from resident GPU sampler paths

Wire the v2 preview callback schedule into the backend-resident sampler loop so Anima/Qwen/Wan/Flux flow paths can emit TAE preview frames without changing final latent residency.

Force preview-only TAEHV contexts to decoder-only loading, handle single-image TAEHV preview outputs, and document Anima/Wan TAEHV preview usage.

Complete SDXL sampler coverage for GPU backend path

Paralol BF16 VAE patchset build 4f99afe

Paralol VAE encode GPU handoff ff9c6d6

Paralol TAE previews build d84733b

Enable Anima VAE encode bridge

Allow Anima to use the existing Wan/Qwen VAE compatibility bridge for modular VAE Encode handoff. The bridge returns a 16-channel GPU latent handle, routes Anima VAE-encoded handles through the Wan/Qwen decode bridge, and reports CPU bridge copies and large IM2COL workspaces honestly. Strict GPU-resident mode still refuses the bridge path.

Document Paralol fork delta