Use local pointer in UTF8StreamJsonParser fraction and long-name loops

[pjfanning]

Not yet ready for merge: needs re-measuring with more forks (results below are noisy) and review.

Applies the hot-loop lesson from #1775 / #1682 to UTF8StreamJsonParser: C2 doesn't keep a field in a register across loop iterations, so a scan loop that advances _inputPtr directly stores it on every byte.

• _parseFloat(): the fraction-digit loop now scans the current buffer with a local pointer (the outer loop still handles _loadMore()).
• parseLongName(): local pointer and local quad buffer; _inputPtr is written back before each exit (parseEscapedName() continues from it).
• _parseUnsignedNumber() / _parseSignedNumber(): deliberately not changed, with a comment. With a local pointer C2 treats the integer loop as a counted loop and unrolls it: 13-digit longs got 10-17% faster, but typical 1-10 digit ints got 6-12% slower on JDK 25 (perfnorm: fewer instructions, but DSB->MITE shift and 2x branch misses; the gap reversed with -XX:LoopUnrollLimit=0).

String value loops (_finishString() etc.) already use local pointers.

Results

ParseBench, nextToken() over a whole document: 2000 ints (1-10 digits), 2000 13-digit longs, 2000 doubles (123.4567891), and 100 objects with 40 property names of 16-40 bytes. i7-8550U/WSL2, 5 forks, us/op:

JDK	doc	3.x	this PR	change
17	doubles	132.36 ± 2.99	125.46 ± 2.60	-5.2%
17	names	782.04 ± 29.35	636.22 ± 40.76	-18.6%
25	doubles	140.59 ± 3.51	119.66 ± 5.25	-14.9%
25	names	709.83 ± 19.74	731.71 ± 36.09	+3.1%

Caveats:

• This benchmark moves between runs. 3.x on JDK 25 ints scored 88-93 in four runs and 73 in one, with all forks agreeing within each run. The ints/longs docs run unchanged code (identical bytecode and inlining decisions, checked with PrintInlining), so they are left out of the table.
• Names on JDK 17 was -4.4% in an earlier run of the same parseLongName() change, so the size of that gain is uncertain. JDK 25 doubles was -14% in both runs.

🤖 Generated with Claude Code

[github-actions]

📈 Overall Code Coverage

Metric	Coverage	Change
Instructions	86.52%	📉 -0.010%
Branches	80.60%	📉 -0.020%

Overall project coverage from JaCoCo test results. Change values compare against the latest base branch build.