Fix numpy 2.x compatibility in face3d utilities

[AbdullahKhan123-tis]

Summary

Three files in src/face3d/ crash or raise warnings when running with numpy 2.x. This PR fixes all of them.

src/face3d/util/preprocess.py

• Wrap
  p.VisibleDeprecationWarning in ry/except AttributeError - this warning class was removed in numpy 2.0 and raises AttributeError on import
• Replace loat((t[0] - w0/2)*s) with
  p.array(...).item() to avoid the numpy 2.0 error "only size-1 arrays can be converted to Python scalars" when [0] is a 1-element array
• Explicitly extract scalar values from [0] and [1] via .flat[0] before passing to
  p.array([w0, h0, s, t[0], t[1]]), fixing the "setting an array element with a sequence" inhomogeneous shape error

src/face3d/util/my_awing_arch.py

• Replace preds.astype(np.float, copy=False) with preds.astype(float, copy=False) -
  p.float was deprecated in numpy 1.20 and removed in numpy 2.0

src/face3d/extract_kp_videos_safe.py

• Add progressive fallback confidence thresholds for detect_faces (0.97 ? 0.5 ? 0.2). The default 0.97 threshold can miss valid faces in portrait photos where the face is slightly off-center or has non-standard framing, causing an IndexError on �boxes[0]. The fallbacks ensure detection succeeds without compromising quality for well-framed inputs.

Test plan

• Tested end-to-end on PyTorch 2.5.1+cu121, numpy 1.26.4, RTX 4050 (6 GB VRAM)
• Landmark detection, 3DMM extraction, and face rendering all complete without error
• No regressions on standard portrait inputs