Spend the maintenance interval between passes, not before the first

[zaeku]

Moves the maintenance.interval sleep from the beginning to the end of the loop body.

maintenance.interval currently sleeps before the first pass, causing newly started processes to sit idle for one interval. Ephemeral maintainers (e.g., serverless hosts or per-request containers) often terminate before executing any work without raising an error.

Moving sleep to the end of the loop fixes this while keeping the drain check at the top.

Trade-off: Simultaneous restarts will no longer be staggered across the interval. Since leases/ guards heavy operations, this results in lock contention rather than duplicate work. I can put this behind a feature flag if default staggering is preferred.

[0bserver07]

Ran this here: cargo test -p walgit-server --test maintain --test drain passes, and the drain check staying at the top keeps D31 intact.

Two things worth stating in the PR, since they're the actual trade-off:

• The first pass now starts before prewarm has warmed anything. run_pass opens with registry.list() and then one unit per assigned repo, and on the host that maintains a large repository that unit can be a full Serve sync or a bundle build, so it now races startup instead of getting one interval of serving-only warmth. Might be fine, might be what the leading sleep was for; the file's history doesn't say (it's in the initial release).
• On simultaneous restarts, compaction and bundle builds are behind leases, but the checkpoint unit isn't (maintain.rs:588), so two hosts would both write the same checkpoint. Idempotent by D22, so wasted round trips rather than duplicate work, but this repo counts round trips.

follow::run_loop has the same sleep-first shape (follow.rs:39), so whichever way this lands, that loop probably wants the same treatment.