Add image complexity filter to prune blank/simple images

[nicpottier]

Summary

• Adds a grayscale standard deviation filter (min_stddev) that prunes visually simple images (blank backgrounds, solid fills) during image classification, running after existing size filters
• Exposes the complexity threshold in the v2 Extract Settings UI under "Image Filters"
• Fixes a brief "Loading pages..." spinner flash when starting a rerun from settings across all v2 step views

Details

Images from PDFs often include blank or near-blank artifacts (solid backgrounds, white boxes) that pass size filters but add no value. This mirrors the is_blank_image approach from adt-press.

Pipeline changes:

• New grayscaleStdDev() function in image-complexity.ts — pure JS decoding via jpeg-js and pngjs
• min_stddev added to ImageFilters Zod schema
• classifyPageImages accepts an optional getImageBytes accessor; complexity check only runs on images that pass size filters
• Default threshold of 2 added to all config presets (storybook, reference, textbook)

UI changes:

• Complexity input added to v2 ExtractSettings under "Image Filters" heading
• Loading guard fix (isLoading && !stepRunning) applied to all 7 v2 step views

Test plan

• New unit tests for grayscaleStdDev (solid PNGs, gradient, JPEG, unsupported format)
• New integration tests for complexity filter in classifyPageImages (prune blank, keep complex, skip when unconfigured, skip when no bytes accessor, skip already-pruned)
• Run pipeline on a PDF with known blank images, verify they're pruned with stddev reason
• Verify no loading spinner flash when starting a rerun from settings