perf: reduce TOML and repeated join allocations

[He-Pin]

Motivation

std.manifestTomlEx spends avoidable time and allocation in TOML string/key rendering and table traversal. The go_suite/manifestTomlEx benchmark is sensitive to regex matches, intermediate escaped strings, Seq path copies, repeated sorting, and per-value renderer allocation.

cpp_suite/large_string_join builds std.makeArray(76846, function(_) 'x') and then joins it. That used to allocate a large backing array filled with the same constant value, then walk it again in std.join.

Modification

• Write TOML strings and keys directly to the StringWriter.
• Replace bare-key regex matching with a small ASCII scan.
• Avoid TOML array-render string concatenations and StringWriter.flush().
• Use the sorted key cache plus section flags instead of partition plus two sorts.
• Reuse one TomlRenderer per table level and keep table paths in a mutable buffer.
• Represent large constant std.makeArray results as a constant array view.
• Teach string std.join to consume constant/repeated string arrays with exact-size StringBuilder output and ASCII-safe result tagging.

Results

JMH regression cases, upstream/master at 8b67cb1e:

benchmark	upstream/master	this PR	delta
go_suite/manifestTomlEx.jsonnet	0.073 ms/op	0.054 ms/op	26.0% faster
cpp_suite/large_string_join.jsonnet	0.625 ms/op	0.461 ms/op	26.2% faster

Native CLI hyperfine against latest source-built jrsonnet master 5b43fa8 (jrsonnet 0.5.0-pre98):

benchmark	upstream/master	this PR	latest jrsonnet	highlight
cpp_suite/large_string_join.jsonnet	5.2 +/- 0.4 ms	4.7 +/- 1.2 ms	8.3 +/- 1.2 ms	this PR is 1.77x faster than latest jrsonnet; CLI timings are short/noisy

Verification

• git diff --check
• ./mill -i __.checkFormat
• ./mill -i 'sjsonnet.jvm[3.3.7].test'
• ./mill -i 'sjsonnet.js[3.3.7].test'
• ./mill -i 'sjsonnet.native[3.3.7].test'
• ./mill -i 'sjsonnet.jvm[2.12.21].compile'
• ./mill -i 'sjsonnet.jvm[2.13.18].compile'
• ./mill -i 'sjsonnet.js[2.13.18].compile'
• ./mill -i 'sjsonnet.native[2.13.18].compile'
• ./mill -i 'sjsonnet.native[3.3.7].nativeLink'
• ./mill -i bench.runRegressions bench/resources/go_suite/foldl.jsonnet bench/resources/cpp_suite/bench.04.jsonnet bench/resources/cpp_suite/large_string_join.jsonnet bench/resources/go_suite/manifestTomlEx.jsonnet
• hyperfine --warmup 10 --min-runs 50

Boundary Checks

• TOML output order remains non-section keys first, then section tables; both groups stay sorted.
• TOML section classification still follows visibleKeyNames order to preserve the previous evaluation/trace order.
• Existing TomlRenderer.escapeKey API is preserved.
• Constant array views materialize back to the same repeated Eval array when an existing caller asks for asLazyArray.
• std.join falls back to the existing element-by-element path for non-string constants, mixed arrays, and lazy values.

References

• Head: a2a96ed47bf87ea33fab7cbc6a47642c358b3e75
• Latest source-built jrsonnet: 5b43fa88b8c43856dd5a2daa9c5c251153c5e14d