ci: parallelize e2e suite + per-test timing

[Sunrisepeak]

Why

The build + test (linux x86_64, self-host) job ran everything in series — build → unit tests → e2e (~18m) → toolchain matrix (3 clean rebuilds) → xlings integration — for a ~34m critical path per PR.

What

1. Split the e2e suite into its own parallel workflow (ci-linux-e2e.yml).

• It shares ci-linux's cache lineage (mcpp sandbox + xlings + target/), so it restores a warm build; the only added cost vs. inline is one warm mcpp build.
• ci-linux.yml drops the e2e step and keeps build + unit/integration + toolchain matrix + the xlings integration build. Its toolchain steps no longer rely on e2e's warm-ups: GCC uses the mcpp.toml-pinned gcc@16.1.0 (already in the sandbox), and the musl --target build auto-installs gcc@15.1.0-musl on demand.
• Per-PR critical path: max(build+matrix, build+e2e) instead of their sum.

2. Per-test timing in tests/e2e/run_all.sh.

• Each test now prints its elapsed time (PASS: 12_foo.sh (3.40s)), via a portable ms timer (bash 5 EPOCHREALTIME, falling back to whole-second date on old bash like macOS /bin/bash 3.2).
• The summary prints a slowest-first table + executed-total, so the long-pole tests surface for future sharding/optimization instead of hiding behind OK.

E2E timing (slowest first; executed total 4m51s):
    1m54s  36_llvm_toolchain.sh
      48s  02_new_build_run.sh
      ...

⚠️ Required-checks action (admin)

The new check e2e suite (linux x86_64, self-host) must be added to main's required-status-checks ruleset, otherwise PRs could merge without e2e passing. The existing build + test (linux x86_64, self-host) check name is unchanged.