Fix image dedup collapsing distinct images that share generic alt text (#286)

[Imccccc]

Problem

_deduplicateImages keyed its out-of-figure pass on alt text: a run of images sharing the same non-empty alt but differing in src was treated as lazy-load / reader-mode duplicates and collapsed to a single image.

That assumption holds for a large/small variant pair, but breaks whenever many distinct images share a generic identical alt:

• image galleries where every image has alt="image" (the case in Lazy-loaded article images with generic identical alt text are deduplicated away #286), and
• WeChat 公众号 (mp.weixin.qq.com) articles, where every inline image is emitted with alt="Image".

On a real WeChat article this collapsed 28 distinct inline images down to 1 — the long-standing "Web Clipper only saves the first image from 公众号" symptom traces directly here.

Root cause

In the second loop of _deduplicateImages, the duplicate signal was same alt && different src. With N distinct images all carrying the same generic alt, each pair matched and all but one were removed.

Fix

Dedup the out-of-figure pass on image identity rather than alt:

• Two images are duplicates only when their normalized src (host + path, ignoring the query string) matches but the raw src differs — i.e. a size/query variant of the same asset.
• Distinct images (different normalized src) are never merged on alt alone.
• The existing identical-src = legitimate repeat behavior is preserved (e.g. the same formula image used twice).

The in-figure pass and the lightbox pass are unchanged.

Testing

• New tests/image-dedup-generic-alt.test.ts: distinct images sharing an identical generic alt / empty alt / WeChat-style mmbiz markup all survive; a legitimately repeated identical-src image still keeps both copies.
• Full suite green (npm test): 312/312, including the existing elements--image-dedup, elements--lazy-image, elements--lightbox-image-dedup, and elements--svg-placeholder-lazy-image fixtures — no regressions.
• Verified on the raw post-scroll HTML of a real WeChat article: 1 → 28 inline images. (Remaining drops are a 1×1 SVG tracking pixel and sub-130px decorative icons, correctly filtered.)

Fixes #286.