Fix Array.__delitem__ corrupting rendered output on slice deletion

[Fries-tempura]

Summary

Array.__delitem__ keeps the in-memory list and the rendered value groups in sync, but for slice deletions it recomputes which rendered positions to drop with:

range(key.start or 0, key.stop or length, key.step or 1)

This mishandles ordinary slices:

• stop == 0 (a[:0]): 0 or length evaluates to length, so a no-op deletion removes every rendered element.
• negative bounds aren't normalized: a[-1:] → range(-1, length), and a[:-1] → range(0, -1) (empty, removes nothing).

list.__delitem__ handles the slice correctly, so the list contents stay right, but as_string() / dumps() silently disagree with them:

import tomlkit
doc = tomlkit.parse("a = [1, 2, 3]\n")
del doc["a"][:0]        # deletes nothing
tomlkit.dumps(doc)      # 'a = []\n'  — expected 'a = [1, 2, 3]\n'

operation	list(arr)	arr.as_string()	expected
del arr[:0]	[1, 2, 3]	[]	[1, 2, 3]
del arr[-1:]	[1, 2]	[]	[1, 2]
del arr[:-1]	[3]	[1, 2, 3]	[3]

Fix: use slice.indices(length), which resolves None, negative, and out-of-range bounds exactly the way list.__delitem__ does, so the removed render positions always match the list.

Tests: added test_array_slice_deletion_keeps_render_in_sync covering a[:0], a[-1:], and a[:-1]. Full tests/test_items.py passes (87 passed).

Agent Drafting Metadata

• Agent: Claude Code (property-based testing workflow)
• Model: Claude Opus 5
• Notes: The bug was found by property-based testing (Hypothesis) run against main. The one-line fix and the regression test were drafted with the agent and verified locally — the repro fails before the change and tests/test_items.py passes (87 passed) after. Reviewed by the submitter before opening.