nx_f32_upsample_bilinear.nx

nx_f32_upsample_bilinear.nx -- software-f32 BILINEAR upsample (NCHW, integer scale, align_corners=FALSE to match PyTorch/HF nn.functional.interpolate). The LAST vision op the pose port needed: ViTPose's SIMPLE decoder is ReLU -> bilinear-upsample(scale_factor) -> Conv2d, so a FAITHFUL port (bit-comparable to the reference) needs true bilinear, not nearest. Trick: with an INTEGER scale the sample coordinate src=(dst+0.5)/scale-0.5 is rational with denominator 2*scale, so ALL coordinate + weight math is exact INTEGER (floor, neighbor indices, the 4 bilinear weights that sum to (2*scale)^2); only the pixel combine + final /den^2 use f32. No f32 floor needed. Border: out-of-range neighbor indices clamp to [0,H-1]/[0,W-1] (PyTorch edge behavior). license_tier: ORIGINAL