testing: establish a load testing environment for realistic scale validation

[savme]

Problem

We have no way to run sustained load tests against Milo at production-scale concurrency. The benchmark suite introduced in #615 runs in a local environment only. This gap has real consequences:

• Optimization work has no feedback loop — we can't tell if a change improves or regresses scaling behavior before it ships
• Per-project resource overhead (storage: project control plane storage layer does not scale with project count #596) can't be measured at the project counts that matter
• Concurrency bugs only manifest under burst concurrency — they're invisible in unit tests and low-traffic staging

What we need

A dedicated environment that can:

• Sustain hundreds to low thousands of projects without becoming unstable
• Run repeatable admission burst scenarios (project re-engagement, concurrent API traffic)
• Expose the same goroutine, memory, and etcd watcher metrics we track in production
• Be torn down and rebuilt cleanly between runs

Out of scope

Full production parity. The goal is a controlled environment that reproduces the scaling behaviors we care about — not a mirror of prod infrastructure.

Success

• The storage: project control plane storage layer does not scale with project count #596 benchmark can run to 1K+ projects without the environment being the limiting factor
• Goroutine and memory regressions are detectable before a PR merges
• Concurrency bugs are reproducible on demand rather than inferred from production profiles