Move hot-path consumers back to direct scans.

[jhy]

Make the parser-hot CharacterReader scanners use direct delimiter loops instead of routing through consumeMatching(CharPredicate).

The lambda predicates looked cleaner, but would not always get a good JIT and could stay in virtual dispatch. That led to a wide variance in throughput.

This keeps the generic predicate scanner for lower-traffic cases, but gives the main HTML tokeniser scanners monomorphic loops that the JIT can compile more predictably.

Fixture	Baseline avg ops/s	Baseline min	Baseline max	Baseline spread	New avg ops/s	New min	New max	New spread
tiny	447,238	375,236	497,448	27.3%	471,773	460,019	478,158	3.8%
small	58,630	58,031	59,325	2.2%	85,310	81,470	86,786	6.2%
medium	20,764	19,626	27,966	40.2%	27,648	27,524	27,836	1.1%
large	597	592	607	2.5%	598	591	606	2.5%
x-large	184	182	186	2.4%	182	175	184	5.2%

The spread really seemed to be dependent on the contents of the fixture more than the size.