perf: speed up index generation sorting

[RomainLanz]

Hey! 👋

VirtualFileSystem.scan currently uses naturalSort, which calls localeCompare with collation options for every comparison. Sorting becomes the main CPU cost when generating indexes with hundreds or thousands of files.

This PR creates one Intl.Collator with the same options and reuses its bound comparator. It preserves the existing natural, case-insensitive, and accent-insensitive order. A focused test covers numbers, leading zeros, casing, accents, and nested directories.

Benchmark

Node.js 24.20.0 using the public IndexGenerator API, with 12 independent before/after process pairs and alternating execution order.

Files	Before	After	Change
10	0.750 ms	0.628 ms	-16.3%
25	1.877 ms	1.435 ms	-23.5%
50	3.266 ms	2.150 ms	-34.2%
100	5.158 ms	2.756 ms	-46.6%
1,000	34.575 ms	10.908 ms	-68.5%
5,000	184.340 ms	51.217 ms	-72.2%

At 1,000 files, the CPU profile dropped from 2,120 ms to 673 ms across 60 generations.

The old and new comparators produced identical results across 4,206,601 comparisons and 100 shuffled stable sorts. A generated 1,000-file index was also byte-identical.

Validation

• 262 tests passed
• Typecheck passed
• ESLint passed