fix(ci): avoid redundant heavy queue benches

[AztecBot]

Related: AztecProtocol/aztec-packages#24454

Summary

• pass BENCH_UPLOAD=0 to every non-primary multi_job_run shard
• make ci-full and ci-full-no-test-cache skip the post-test bench when BENCH_UPLOAD=0
• keep ordinary single-shard full CI benchmarking unchanged via the default BENCH_UPLOAD=1

Investigation

The dequeued merge-group run did not fail from a merge conflict. PR #24454 is mergeable and its head CI passed. The failed merge-group run had tests complete successfully on the failing full shards, then both ci/x7-full and ci/x9-full entered bench all before being marked failed. The later merge-group run similarly had all observed full shards pass tests, with a remaining shard still spending time in bench all.

This change keeps one benchmark artifact producer and removes the redundant post-test benchmark sweeps from the other heavy queue shards.

Testing

• bash -n ci.sh && bash -n bootstrap.sh && git diff --check
• BENCH_UPLOAD=0 gate check: bench_if_enabled skips, while the default path calls bench
• attempted ./bootstrap.sh ci, but this ref does not define a ci target and exits with Unknown command: ci
• shellcheck ci.sh bootstrap.sh unavailable: shellcheck is not installed in this container

---

Created by claudebox · group: slackbot

[AztecBot]

Automatically closing this stale claudebox draft PR (no updates for 5+ days). Re-open if still needed.