Fix subpixel misalignment when an image dimension is not a multiple of 4

[f-dy]

Closes #18

Keypoint coordinates are misaligned by up to half a pixel when an image dimension is not a multiple of 4.

Cause

forward_impl upsamples the logits back through the sizes recorded by the encoder. Those are floor-halvings, so a dimension that does not halve exactly makes an F.interpolate ratio non-integral, and with align_corners=False the score map lands off the pixel grid.

Only the two finest levels move a keypoint measurably, so the condition is % 4, not % 8:

width	%8	%4	upsample ratios (fine to coarse)	biased
497	1	1	2.0040, 2.0000, 2.0000	yes
498	2	2	2.0000, 2.0081, 2.0000	yes
499	3	3	2.0040, 2.0081, 2.0000	yes
500	4	0	2.0000, 2.0000, 2.0161	no
508	4	0	2.0000, 2.0000, 2.0161	no
504	0	0	2.0000, 2.0000, 2.0000	no

500 and 508 are the discriminating cases: not multiples of 8, but unbiased, because only the coarsest ratio is non-integral and that level is smoothed away.

load_image floors both dimensions with (scale * W) // 8 * 8. A multiple of 8 is also a multiple of 4, so that path satisfies the condition and never exposes this. A caller passing its own tensor to detect gets no such guarantee.

Fix

Pad the bottom-right up to a multiple of 4, then crop the score map back. Padding after the origin leaves the top-left alignment unchanged, so sizes that were already correct produce identical output.

Padding to 8 instead would also make the coarsest halving exact. Measured against padding to 4, that moves keypoints by 0.0025 px and is bit-identical wherever both round to the same size, so 4 is used as the smallest sufficient modulus. Overhead is under 0.2% of a detect call (0.04 ms of ~23 ms on a T4).

Measured

Trimmed-mean flip-equivariance residual, which must be 0. to_pixel_coords maps index i to i + 0.5 on a [0, W] axis, so a flip maps x to W - x exactly. One image cropped to each size, so content is constant.

dim % 4	x before	x after	y before	y after
0	+0.001	+0.008	-0.004	-0.003
1	-0.430	+0.014	-0.506	-0.002
2	-0.152	+0.007	-0.191	-0.004
3	-0.570	+0.008	-0.612	-0.002

x follows W % 4 and y follows H % 4 independently: across 16 widths the y residual stays near -0.002, and across 16 heights the x residual stays near +0.010.

More keypoints also survive the flip, since they no longer move off their true position (width 503: 929 to 1191 correspondences).

Tests

tests/flip_equivariance_test.py covers one size per residue class on both axes. Without the fix 12 of 16 cases fail and the 4 that pass are exactly the % 4 == 0 sizes, so it is not vacuous. With the fix all 16 pass, and tests/smoke_test.py is unchanged at 1 passed.

Found while measuring keypoint coordinate conventions through LoMa, which carries a copy of this detector and shows the same behaviour; davnords/LoMa#14 is the same patch there.

Draft because you may prefer the padding in detect rather than forward_impl.

[f-dy]

Superseded by #19, which has the same change with a clean history.