perf: speed up TOML manifestation

[He-Pin]

Motivation:

std.manifestTomlEx spends avoidable time and allocation while rendering TOML: regex matching for bare keys, intermediate escaped strings, per-value renderer allocation, Seq path copies, partition, and repeated sorting during table traversal.

Key Design Decision

Keep this PR TOML-only. The earlier repeated-join work is now handled separately in #825, so this branch was intentionally reduced to a single focused TOML rendering commit after rebasing onto latest master.

Modification:

• Write TOML strings and keys directly to the existing StringWriter.
• Replace bare-key regex matching with a small ASCII scanner.
• Avoid string concatenations and unnecessary StringWriter.flush() calls while rendering arrays.
• Reuse a TomlRenderer per table level and keep table paths in a mutable buffer.
• Use sorted visible keys plus section flags instead of partition plus extra sorting.
• Add directional golden coverage for TOML key escaping, non-section ordering, nested tables, and table arrays.

Benchmark Results:

JMH on JDK 21, ./mill -j 1 bench.runRegressions bench/resources/go_suite/manifestTomlEx.jsonnet, lower ms/op is better; ops/ms = 1 / ms_per_op, higher is better.

Case	master ms/op	PR ms/op	master ops/ms	PR ops/ms	Delta
go_suite/manifestTomlEx.jsonnet	0.067	0.053	14.93	18.87	+26.4% ops/ms

Scala Native hyperfine, lower is better, comparing locally built Scala Native binaries and source-built jrsonnet 0.5.0-pre98.

Case	master	PR	jrsonnet	Result
go_suite/manifestTomlEx.jsonnet	5.3 ± 1.3 ms	5.3 ± 1.3 ms	4.0 ± 1.1 ms	startup/noise dominated; PR and master are neutral
scaled TOML manifestation loop	29.0 ± 1.6 ms	12.1 ± 1.0 ms	10.0 ± 0.9 ms	PR is ~58.3% faster than master; still ~1.21x jrsonnet

Analysis:

The output remains compatible with current sjsonnet behavior: local master and this PR produced identical output on the benchmark inputs and the new directional TOML golden test. The scaled Native comparison includes jrsonnet as a performance reference, but jrsonnet's TOML formatting differs from sjsonnet on these inputs, so sjsonnet-vs-sjsonnet output equality is the semantic gate.

References:

• PR: perf: speed up TOML manifestation #828
• Rebased head: 3efa453a5b6f48d6dfd0e1cd034922d6c8e2ff03
• Base: 0ae7b78a93c4e643f9bcfb6dd1d99d9fe7a522a9
• Split dependency: perf: speed up constant string joins #825 now owns the repeated constant join optimization.

Result:

• ./mill --no-server -j 1 __.reformat passed locally.
• ./mill --no-server -j 1 __.test passed locally.
• PR comments were reviewed; there are no review threads. The maintainer rebase request is addressed by this force-pushed TOML-only branch.

[stephenamar-db]

rebase

[He-Pin]

Rebased onto latest master and reduced this PR to TOML-only so it no longer duplicates #825's repeated-join work. I refreshed the local validation and benchmark data in the PR description.