Add side-by-side map and cutout grid to --compare-residuals

[Athanaseus]

Numbers alone can't show what changed between two images or where. --compare-residuals now also writes SourceResidualCutouts.html (with -catalog) or RandomResidualCutouts.html (without): a side-by-side map, then a grid of the 18 biggest changes as stamp pairs. The existing noise-ratio plot is unchanged. Image-agnostic: two self-cal rounds, DI vs DD, anything on the same pixel grid.

Side-by-side map

• Both images downsampled for display, each marking in red only where it is deeper than the other image — marking every negative puts the same clumps on both panels, since negatives cluster around the same bright sources.
• Marks come from block minima, not means, so downsampling cannot erase a hole.
• One threshold for both, 5× the noisier image's MAD noise, so the cleaner image can't pick a threshold that flatters it. Each title gives that image's pixel count below it: here 36,405 vs 58,066.
• The 18 grid positions are circled and labelled; hover gives the values; zooming one panel zooms the other.

Cutout grid — with a catalogue

• Full-resolution 64×64 stamps (never downsampled: a negative bowl is a few beams across), image 1 and 2 touching, one colour scale per pair.
• Picks the sources whose deepest negative changed most, in either direction, so image order doesn't matter — swapping the images picks the same sources and flips which one is reported worse. Overlapping stamps are skipped, so a close pair doesn't show the same bowl twice.
• Here: SRC1523 is the over-subtracted peeled source; SRC739 nearly vanishes; most others show new ripple structure in image 2.
• --reference-position adds one extra pair at that position, on its own row above the grid (top row here).

Cutout grid — without a catalogue

• The 18 biggest pixel changes anywhere (image 2 − image 1), either direction. This answers "where did the sky change most" rather than "where did new artefacts appear", because a few mJy of flux change outranks a sub-mJy ripple.
• Here the top entries are the bright sources the DD run peeled (its output is named brighptsrc5). Bright sources saturate on the negative-driven colour scale, so the size of each drop is in the title.

Options

• -cr: everything in one tabbed ResidualReport.html (Noise ratio · Cutouts). -cr previously covered only catalogue comparisons.
• -svg: also writes the map (…_map.svg) and grid (….svg) through matplotlib, which is optional; a warning if it's missing.
• --reference-position is checked before any work: a position off either image or on blanked pixels raises immediately, as it already does for dynamic range. That check is now one shared helper.

Also fixed

• The CLI never passed -svg or -cr to --compare-residuals.
• Once -svg reached it, bokeh's SVG export of the noise-ratio plot aborted the run on machines without a browser driver. It now warns and continues; the matplotlib SVGs need no browser.
• --reference-position help didn't mention that sexagesimal is accepted.

Notes

• Both images must share a pixel grid (size, projection, reference pixel, pixel scale). If they don't, the cutouts are skipped with a warning and the rest of the run goes ahead.
• Per-source rows gain ra_deg and dec_deg, appended at the end, so stamps don't rely on catalogue names being unique.
• On two 6075² images: ~30 s, ~800 MB peak memory, ~6.5 MB HTML (mostly the two map images).