perf: use ASCII bitsets for strip chars

[He-Pin]

Motivation:

std.stripChars, std.lstripChars, and std.rstripChars already avoid boxed Set[Int] for single/BMP delimiter sets on current master. This draft keeps that Unicode-aware path and narrows the ASCII-only multi-character delimiter case by replacing the per-call java.util.BitSet allocation with primitive masks.

Key Design Decision:

• ASCII-only delimiter sets use primitive masks before the BMP fallback.
• JVM and Scala.js use four Int masks; Scala Native uses two Long masks.
• Non-ASCII BMP delimiter sets keep the upstream BitSet fast path.
• Surrogate-containing delimiter sets still fall back to codepoint iteration.
• chars is forced before str, preserving existing error evaluation order.

Modification:

• Add platform-specific ASCII strip mask helpers.
• Route stripChars, lstripChars, and rstripChars through the ASCII mask path when chars contains only ASCII delimiters.
• Keep existing Unicode/BMP/codepoint behavior unchanged.
• Extend strip tests for ASCII, Unicode fallback, mask boundaries, and evaluation order.

Benchmark Results:

Rebased onto current upstream/master at ca61a7a391e0a6910fa51a8891c69e86d708227e.

JMH command:

./mill --no-server -j 1 bench.runRegressions \
  bench/resources/go_suite/lstripChars.jsonnet \
  bench/resources/go_suite/rstripChars.jsonnet \
  bench/resources/go_suite/stripChars.jsonnet \
  bench/resources/cpp_suite/bench.09.jsonnet

Case	master ms/op	PR ms/op	Result
lstripChars.jsonnet	0.113 / 0.115	0.115 / 0.113	mixed/no stable delta
rstripChars.jsonnet	0.116 / 0.116	0.121 / 0.115	mixed/no stable delta
stripChars.jsonnet	0.112 / 0.119	0.115 / 0.117	mixed/no stable delta
bench.09.jsonnet	0.041 / 0.042	0.042 / 0.042	neutral

Scala Native hyperfine, original go_suite strip files, 50 runs, compared against source-built jrsonnet 0.5.0-pre98:

Case	master	PR	jrsonnet	Result
lstripChars.jsonnet	6.7 +/- 2.4 ms	2.7 +/- 1.3 ms	1.1 +/- 1.0 ms	too short/noisy, PR faster in this run
rstripChars.jsonnet	5.7 +/- 1.2 ms	3.2 +/- 0.6 ms	4.8 +/- 2.1 ms	too short/noisy, PR faster in this run
stripChars.jsonnet	6.0 +/- 2.3 ms	3.9 +/- 1.2 ms	3.4 +/- 1.7 ms	too short/noisy, PR faster in this run

Scaled temporary Native strip stress cases, 30 runs, --shell=none, outputs checked equal across master/PR/jrsonnet:

Case	master	PR	jrsonnet	Result
lstrip-scaled.jsonnet	8.0 +/- 3.0 ms	7.1 +/- 0.5 ms	143.9 +/- 3.6 ms	small PR improvement
rstrip-scaled.jsonnet	7.9 +/- 0.6 ms	7.8 +/- 0.5 ms	143.2 +/- 7.0 ms	neutral/slightly positive
strip-scaled.jsonnet	7.6 +/- 0.7 ms	7.3 +/- 0.9 ms	142.6 +/- 4.3 ms	small PR improvement

Analysis:

Current master already has the main BMP BitSet strip fast path, so this PR no longer closes a visible go_suite strip gap on JVM JMH. The remaining value is narrower: a GC-friendly ASCII delimiter cleanup that can be slightly positive in scaled Native strip stress cases, while original go_suite cases are too small/noisy to justify moving this out of draft.

References:

• Source/provenance: He-Pin/sjsonnet@dd90b11
• Current base: ca61a7a

Result:

Keep as draft. This is safe after rebase and tests, but it should not be marked ready unless a future benchmark shows a stable user-visible improvement or we explicitly decide the GC/allocation cleanup is worth merging on its own.

Validation:

• Rebased branch head: 17bee2380f5b1df02a9f975ccb7807a71241f39d
• Local validation after rebase: ./mill --no-server --ticker false --color false -j 1 __.test -> Tests: 441, Passed: 441, Failed: 0

[He-Pin]

Closing as low priority after rebase: current master already has the main strip-char fast path, and this PR shows no stable standout delta on the go_suite strip docs cases or Native hyperfine.

[He-Pin]

Reopened. This is not negative; keep it as draft as a small GC-friendly strip delimiter cleanup. Current docs strip benchmarks do not show standout gains, so it should not be moved to ready yet.

[He-Pin]

Rebased to latest upstream/master (0ae7b78a). Full __.reformat + __.test passed locally. Fresh JMH is mixed/no stable positive signal; Native scaled strip stress cases are slightly positive, so I kept this as draft rather than moving it ready. PR body has the updated benchmark table and current decision.

[He-Pin]

Rebased to current upstream/master (ca61a7a) and force-pushed with lease: 7f0efb1 -> 17bee23. Local validation: ./mill --no-server --ticker false --color false -j 1 __.test passed (Tests: 441, Passed: 441, Failed: 0). Keeping draft because benchmark signal remains mixed.

[He-Pin]

Rebased to cedc083b (master). All 441 JVM tests pass.

Hyperfine go_suite strip benchmarks (warmup 5, 20 runs) on JDK 21:

Case	master ms	branch ms	Result
stripChars.jsonnet	215.3 ± 1.4	546.8 ± 10.7	2.54× slower
lstripChars.jsonnet	213.0 ± 1.0	530.9 ± 9.8	2.49× slower
rstripChars.jsonnet	213.4 ± 1.7	534.6 ± 7.6	2.51× slower

Master already has the BMP BitSet fast path from #748 plus #810/#834 follow-ups, which together beat the 4-Int ASCII mask path here. The split per-codepoint bitmap branching (< 32 / < 64 / < 96 / < 128) appears to defeat what JIT does well on BitSet.get.

Recommendation: close as obsolete. The optimization this PR aimed to introduce has been overtaken by master's strip path.