[Bench] Fixes for pure Fizz bench - #96771

## Summary

Harness fixes, all surfaced while benching a Fizz-only React PR
(react/react#37022):

- **Profile pass now alternates arm order by VM index.** The timed runs
are ABBA, but the CPU profile pass ran base-then-cand on every VM. The
second arm runs warmer (page cache, governor), which deflated every
function in the cand profile by a uniform few percent — untouched Flight
frames "improved" ~7%. With balanced order the same untouched frames
diff to ~1.5% across 14 paired VMs while real movers keep their
magnitude. Each VM records its capture order in the profiles tarball
(`prof-order.txt`) so analyses can check order drift directly.
- **The fingerprint now covers every server-side React layer.**
`FP_FILES` hashed only the vendored Flight server files, which a
Fizz-only React change leaves byte-identical — both arms of a valid A/B
reported the same fingerprint. Added the vendored Fizz bundle and the
shared react-server runtime (hooks/cache) in both suites. The VM boot
loop also warns loudly when arms fingerprint identically, instead of
leaving it for the analysis footer.
- **`--prepare` writes a terminal status phase.** It previously exited
leaving `phase: "measuring"`, so `bench-status` prescribed collecting
measurement data that never existed. `bench-status` now reports
prepare-only runs as complete rather than "DEAD — relaunch".

## Verification

- Full 16-VM e2e run with the ordering + Fizz-fingerprint changes
(`run-pr37022-e2e-prof2`): arms fingerprint distinctly (`18ea275e` vs
`f9805be5`), profile order alternates per VM, untouched hot frames diff
to ~zero in the cross-VM profile aggregate.
- `bench-status.mjs` against a real `--prepare` run dir reports
`complete (--prepare: caches only)`.
- `--dry-run` smoke of the modified launcher; `node --check` on all
three scripts.

<!-- NEXT_JS_LLM -->