fix: keep the plateau rule and the full-resolution fine sweep in SkewCorrector

[OligerMan]

After #20 a document with short lines comes back rotated by one degree: dedoc's
test_module_skew_corrector.test_documents_with_short_lines fails on all five
short_lines-*.png images (0 != 1.0). The failing test is the visible part — on dedoc's test
images the angle changed on 9 of 31, bold_font.png, orient_4.png, scan_orient_1.jpg and
example_with_table5.png among them.

Two things the previous implementation relied on were lost, and it takes both of them to get the
old answers back.

The plateau rule. The projection score of a document with short lines has a plateau rather
than a peak. The old sweep resolved it: if the score two steps away is within 2% of the maximum,
the angle between the two wins. Restored on the fine sweep.

The fine sweep ran on a downscaled image. 1000px on the long side is enough to flip a
borderline one-degree decision. It runs on the thresholded image again; the coarse guess keeps its
512px thumbnail, which is where the speed comes from.

Measured

Angles compared against 0.3.8 on the 31 images of dedoc's tests/data:

build	same angle as 0.3.8	short_lines-*.png	time, all 31
0.3.8	31 / 31	5 / 5	22.5 s
0.3.10	22 / 31	0 / 5	2.6 s
0.3.10 + plateau rule only	29 / 31	5 / 5	2.6 s
0.3.10 + full-resolution fine sweep only	27 / 31	2 / 5	3.7 s
this PR (both)	31 / 31	5 / 5	3.5 s

So the corrector agrees with 0.3.8 on every image again, and stays 6.4x faster than it
(~0.11 s per page against ~1.0 s; 0.3.10 was ~0.06 s).

tests/unit_tests/test_preprocessors.py passes unchanged, flake8 . is clean, and dedoc's
test_module_skew_corrector passes against this build.