perf: reduce allocations

[dsherret]

Reduces allocations when generating print items and when sorting a package.json, without changing output.

• push_str instead of push_string(x.to_string()) for strings, property names and numbers, so no throwaway String is created before dprint-core copies it into its bump allocator.
• true, false, null and , use sc! string containers, so their width is computed at compile time and nothing is allocated per value.
• Comma separated values take an ExactSizeIterator instead of collecting a Vec<Option<Node>> for every array and object. The None case was never used.
• The ignore comment check reuses the leading comments lookup, and dangling comment handling returns early before its line lookup when there are no comments.
• package.json sorting uses PropertySort::by_key, which caches keys, instead of comparators that decoded both property names into Strings on every comparison.
• The .jsonc extension check uses eq_ignore_ascii_case instead of allocating a lowercased string.

Benchmark

Median of 15 runs, comparing main against this branch; the output is asserted identical.

Input	Time	Allocations
~4.7 MB jsonc	894ms → 885ms (within noise)	2,470,190 → 1,490,174 (-39.7%)
package.json with 3000 dependencies	6.18ms → 4.66ms (-24.6%)	78,394 → 69,368 (-11.5%)

On the large file most of the time is spent in dprint-core's printer (~65%) and in parsing, so the allocation drop doesn't show up as a measurable speedup there.