Compare header names case-insensitively without allocating

[hellerve]

header-lookup and header-values lower-cased every request header name on every call: String.ascii-to-lower on each key, so two allocations per header per lookup. Precondition.evaluate makes up to four lookups, and Request.header and if-range-matches? make more, so a plain GET paid this many times over.

This compares names case-insensitively in place instead: lengths first, then byte by byte with A–Z folded to a–z, the same folding tolower does in the C locale. It allocates nothing.

Behaviour is unchanged. Old against new, header-lookup and header-values both, gave 0 divergences over 135,137 queries. The header maps were built from every name of up to three bytes over a, A, z, Z, -, @, [, ` and {, which are the bytes on either side of both letter ranges, plus two names with bytes above 127. The check catches a planted bug at either end of the folding range: not folding Z gives 707 divergences, folding [ gives 976.

Timing. Median of three runs, 200,000 calls each, on an M1 Max, against the headers of a typical browser GET (nine headers):

case	level	master	this branch	speedup
header-lookup, header absent	-O0	1,553 ns	542 ns	2.9×
	-O1	576	52	11.1×
	-O2	574	50	11.6×
	-O3	548	27	20.4×
	-Os	605	84	7.2×
	--optimize	554	28	19.4×
header-values, header present	-O0	1,736	1,221	1.4×
	-O1	658	155	4.2×
	-O2	650	156	4.2×
	-O3	640	156	4.1×
	-Os	674	180	3.7×
	--optimize	644	154	4.2×

Through web 0.12.0's web-build-response, from -O1 up, a build with this change and carpentry-org/time#33's strptime change is 0.4–0.5 µs faster on a plain GET and 0.1–0.2 µs faster on an If-Modified-Since request than one without them. The two changes were not measured apart there.

Tests. Run with -Werror: 516 passed, 0 failed. No test changed. angler and carp-fmt --check, both built from their current HEADs, are clean. carp -x gendocs.carp succeeds.

Request.parse still lower-cases each header name once, to find Cookie. It is not a lookup and is left alone here.