Flaky wall-clock scaling tests can fail CI and the release gate

[rahlk]

Problem

TestTemplateDenseSingleLineParsingScalesNearLinearly (and its sibling TestTemplateParsingScalesNearLinearly) guard against an O(n²) regression by comparing wall-clock time of a 4× larger input against small*9 + 50ms (small*8 + 25ms). On a shared GitHub runner the small measurement lands near 12 ms, so the bound is ~160 ms and ordinary scheduler/GC noise on the large run trips it. Seen on PR #6 (ci job "full suite with a graph", run 34002119560): 8k dense same-line actions took 166ms after 2k took 12ms; want <= 158ms. PR #6 changes no Go code.

Because release.yml runs make test as its hard gate and deletes the tag on failure, this flake can abort a release.

Scope boundary

In: make both scaling tests robust to runner noise without weakening the quadratic-regression guard: larger inputs so the small measurement is well above the noise floor (target ≥ 50 ms on a runner), best-of-N kept, ratio bound kept, absolute ceiling kept.
Out: any change to the parser or to what the tests assert about spans/counts.

Goals

• go test ./internal/dialects/helm -run ScalesNearLinearly -count=30 passes locally and the tests still fail on an injected quadratic regression (temporarily, to prove the guard).
• Total added test time under one second.

Caveats and known risks

• A too-large input hides a quadratic regression behind the absolute ceiling; keep the ratio check primary.

Definition of done

• Both tests pass 30× locally and on CI; the regression guard is demonstrated in the PR description.