perf: batch elements into shared IntersectionObserver instances per threshold

[andreas-fahrecker]

Summary

Groups observed elements by their resolved threshold value and shares
one IntersectionObserver per threshold, instead of creating a new
IntersectionObserver for every observed element.

Motivation

observe() currently creates one IntersectionObserver per element.
On a page with many intersect: elements sharing the same threshold
(the common case, since most elements just use the default threshold
or intersect-half/intersect-full), this means far more observer
instances than necessary. Since restart() disconnects and rebuilds
all observers from scratch, this cost is also repeated on every call
to it (e.g. on SPA route changes).

Changes

• observe() now groups elements by their resolved threshold value
  first, then creates a single IntersectionObserver per distinct
  threshold, observing all elements that share it.
• The intersection callback now reads the target element from
  entry.target instead of closing over the per-element element
  variable, since a single callback now handles multiple elements.
• intersect-once elements are removed from their observer with
  observer.unobserve(el) instead of observer.disconnect(), since
  disconnecting would now stop observing every other element sharing
  that observer.

Backwards compatibility

No API changes. Behavior of intersect-once, intersect-half,
intersect-full, and restart() is unchanged from the outside —
this is an internal implementation change only.

Testing

Tested locally in an Astro project (not the repo's demo page), with
both this change and the arbitrary-threshold PR applied together:

• intersect-once still stops observing the correct element once
  triggered, without affecting other elements sharing the same
  observer/threshold.
• Elements without intersect-once continue to toggle in and out as
  expected on repeated scroll in/out.
• Existing intersect-half and intersect-full still resolve and
  trigger correctly.
• Arbitrary intersect-[value] thresholds (from the other PR) also
  still trigger correctly when combined with this change.

[heidkaemper]

Thank your for the contribution!
I'm going to close this one for the following reasons:

1. The complexity isn't worth it. IntersectionObserver instances are quite cheap. And grouping elements into a Map adds complexity to what is currently a straightforward loop. I'm trying to keep this plugin as small and readable as possible.

2. There seems to be a bug in the element lookup. Using document.querySelector (singular) instead of querySelectorAll caused some problems.

If you ever run into a measurable performance problem here, I'd be happy to look at it again. Thanks again for taking the time.