Fix sync test server port allocation races

[jubishop]

The sync end-to-end harness releases its selected port before td-sync binds it. Another process can take that port, causing a flaky startup failure or letting an unrelated HTTP server's health response make startup appear successful.

Initial setup now retries up to five times, only for the current child's confirmed address-in-use failure. Each failed child is reaped before retry, and the final URL is selected before actors save their configuration. Startup requires the child's post-bind startup message and a successful health check. Restarts retain the existing address and data.

Changes are limited to the Go end-to-end harness and six regression cases covering occupied-port health responses, collision recovery and fixed-address restarts, bounded retries, unrelated failures and stale logs, timeout cleanup, and failed process launch.

Fixes jubishop#2.

Submitted upstream from jubishop#3 with the same commit, ae81b34086de9c1b34e43b7100e648f6dfbf170d.

Validation

Validation recorded in the source PR:

• Red-green regression: the old harness accepted another server's HTTP 200 response on an occupied address; the fixed harness reports the child's bind failure.
• go test -race ./test/e2e -run '^TestServerStartup$' -count=10 passed.
• make test passed, including the complete sync end-to-end suite.
• make lint, GOWORK=off go build ./..., formatting, and whitespace checks passed.
• Test commands used the repository's release-safe feature environment.

The source PR's GitHub Actions cross-build, lint, and test checks also passed for this commit.