Release 0.9.6

[dylanuys]

Relabel EDTalk and FLOAT from semisynthetic to synthetic, and remove VideoUFO from the active registry. The taxonomy treats fully synthesized output as synthetic even when a captured image conditions generation.

Supporting excerpts:

• EDTalk: “fed into the Generator G to synthesize the final result” — §3.1. Identity features condition the generated image.
• FLOAT: “FLOAT can generate a talking portrait video from a single source image and audio” — Figure 1 / §4. Generated motion and source identity latents are decoded into video frames.
• VideoUFO: “we use the official YouTube API to search for videos” — §3, dataset card. Its collection pipeline does not establish uniform visual provenance, so exclude it rather than assign all clips real or synthetic.

Validation: loaded the bundled registry and confirmed both labels and VideoUFO's absence; git diff --check passed. Portrait-method evidence does not independently verify every file in the 34data mirrors.