nx_f32_maxpool2d.nx

nx_f32_maxpool2d.nx -- software-f32 2D max-pooling (NCHW), a STANDARD vision op the f32 tower lacked. Almost every CNN stem (ResNet/HRNet pose backbones) does a 3x3 stride-2 maxpool after the first conv; the pose port needs it. Padding positions are EXCLUDED from the max (standard maxpool: pad with -inf, i.e. never selected). Composes only nx_f32_gt (IEEE-754 magnitude compare), so it is correct for any f32 (unlike raw bit compare on negatives). license_tier: ORIGINAL