chore(ci): apply deploy-then-wait refactor to spartan nightly benchmarks

[AztecBot]

Summary

Applies the same restructuring introduced for the weekly proving benchmark in #23395 to the remaining spartan nightly benchmark workflows, to cut the SSM Undeliverable failures we keep hitting (e.g. the nightly 10 TPS run that died mid network_deploy).

The core idea (unchanged from #23395): stop doing the network deploy + multi-epoch chain warmup inside the single long-running EC2/SSM benchmark command. Instead:

1. deploy-network.yml (runs on the GitHub runner via GKE, no SSM) deploys the network.
2. A wait-for-l2-block job polls until the chain produces its first L2 block.
3. The benchmark then runs with SKIP_NETWORK_DEPLOY=1, so the SSM command only builds + measures against the already-live network — a much shorter SSM lifetime, which is where Undeliverable was striking.
4. A separate cleanup job (if: always()) tears the network down, and notify-failure keeps the existing Slack/ClaudeBox alerts.

Changes

Workflows restructured into select-image → deploy-network → wait-for-l2-block → benchmark (SKIP_NETWORK_DEPLOY=1) → cleanup → notify:

• .github/workflows/nightly-bench-10tps.yml — single 10 TPS benchmark (env/namespace bench-10tps).
• .github/workflows/nightly-spartan-bench.yml — all three benches: benchmark (tps-scenario → ns nightly-bench), proving-benchmark (prove-n-tps-fake), block-capacity-benchmark (block-capacity → ns nightly-block-capacity). The status aggregator and per-bench Slack messages are preserved. All jobs check out ref: next.

SKIP_NETWORK_DEPLOY plumbing (mirrors the proving-bench path from #23395) added to:

• ci.sh: network-bench, network-block-capacity-bench, network-bench-10tps.
• bootstrap.sh: ci-network-bench, ci-network-block-capacity-bench, ci-network-bench-10tps.

Bundled from #23395 (not yet on next)

This PR targets next, but the benchmark changes depend on helpers that currently only exist on merge-train/spartan via #23395. To make this PR self-contained and mergeable to next on its own, it also carries #23395's helper diff:

• deploy-network.yml — notify_on_failure input.
• spartan/bootstrap.sh — wait_for_l2_block command.
• spartan/scripts/wait_for_l2_block.sh — RANDAO-aware wait logic.
• bootstrap.sh / ci.sh — SKIP_NETWORK_DEPLOY for the proving-bench path.
• weekly-proving-bench.yml — the original restructure from chore: refactor weekly proving test wait #23395.

These hunks are identical to #23395, so git will reconcile them when #23395 lands independently. If you'd rather land #23395 first and have me drop the bundled commit, say so.

Note on namespaces

For jobs whose deploy namespace differs from the env-file name (nightly-bench, nightly-block-capacity), the wait step sets NAMESPACE explicitly so wait_for_l2_block polls the right pods (env files use ${NAMESPACE:-default}, so the override is respected).

Not in scope here

The genuine nightly scenario tests run from ci3.yml's ci-network-scenario job (tag-gated), not from test-network-scenarios.yml (a manual workflow_dispatch runner). ci-network-scenario has the same deploy-inside-SSM shape and could get the same treatment as a follow-up.

Test plan

These can't be validated from a sandbox (need AWS/GCP secrets + a real deploy). Before relying on the schedule, run each via workflow_dispatch once:

• Nightly Bench 10 TPS — confirm deploy-network + wait-for-l2-block succeed, then the benchmark runs against the live network with SKIP_NETWORK_DEPLOY=1, and cleanup tears down.
• Nightly Spartan Benchmarks — confirm all three deploy/wait/benchmark/cleanup chains run in parallel without namespace collisions and status reports correctly.

[alexghr]

lgtm