perf(runner): remaining run_test hot-path forks (labels, clock capture, parallel::is_enabled)

[Chemaclass]

Part of #761. Follow-up to #764 (PR #770), which removed most run_test subshells; these are the remaining ones.

Problem

Per test, bashunit::runner::run_test and the surrounding loop still fork:

1. Two normalize_test_function_name captures: label="$(bashunit::helper::normalize_test_function_name ...)" (src/runner.sh:1183) runs on every test; failure_label="$(...)" (src/runner.sh:1191) additionally on hook failures.
2. Clock captures: start_time=$(bashunit::clock::now) (src/runner.sh:1095) and end_time=$(...) (src/runner.sh:1125) whenever durations are measured (the default: needs_test_duration is true because show_execution_time defaults on). On Bash 5 the shell impl (src/clock.sh:127-141) is pure bash yet still pays the capture fork, and it nests a second $(bashunit::clock::shell_time) inside.
3. bashunit::parallel::is_enabled recomputed 3x per test iteration (src/runner.sh:782,789,819; also :830,880 per file). Each call (src/parallel.sh:146) re-evaluates env + OS checks that are constant for the whole run.

Task — checklist, one commit each

• Add a return-slot variant of normalize_test_function_name (src/helpers.sh:40; slot e.g. _BASHUNIT_NORMALIZED_FN_OUT) and convert src/runner.sh:1183,1191. Keep the echoing function for other callers.
• Add bashunit::clock::now_to_slot writing _BASHUNIT_CLOCK_NOW_OUT: the shell branch computes fork-free from EPOCHREALTIME (fold in shell_time); date-seconds computes with one fork instead of two; interpreter/date branches may keep an internal $(...). Convert both run_test call sites. This must not change _choose_impl selection or the perf(clock): skip per-test clock reads when the clock impl forks an interpreter #765 auto-skip behavior.
• Resolve parallel mode once after arg parsing (e.g. _BASHUNIT_PARALLEL_ENABLED set where env layer finishes, after --parallel/--no-parallel are known) and make bashunit::parallel::is_enabled a pure-global read. Beware: keep the internal_log out of the hot reader or early-return it.
• Optional, only if trivial: fold the grep -q "^# bashunit: no-parallel-tests" per-file probe (src/runner.sh:769) into the existing per-file awk scan (build_provider_map, src/helpers.sh:327) so one pass answers both questions.

Tests first (TDD)

Behavior must not change. For the clock slot, add unit tests around now_to_slot for the shell and date-seconds branches (mock EPOCHREALTIME presence). For each return-slot helper add the dynamic-scoping regression test per .claude/rules/bash-style.md (bug precedent PR #672).

Constraints

• Bash 3.0+ only; EPOCHREALTIME fast path is feature-gated exactly as clock::_choose_impl already does.
• Console output byte-identical; durations still reported when enabled.

Acceptance criteria

• ./bashunit tests/ and ./bashunit --parallel tests/ pass after every commit.
• make sa, make lint pass; shfmt -w . produces no diff.
• Report before/after 100-noop timing in the PR (current baseline: ~0.68s / 100 trivial tests on macOS bash 3.2).

Line references valid at commit 1ae3e82.