exec-server: simplify pending environment readiness

[pakrym-oai]

Why

Environment providers need to register an environment before its stable exec-server URL is available. Completing that registration through a later manager lookup by environment ID couples provisioning to mutable registry state: a stale completion can target a replacement that reused the same ID.

This change gives each pending registration its own one-shot completion capability. The connection path only consumes the terminal result associated with its captured environment, which keeps replacement semantics isolated and makes the readiness flow linear.

What changed

• Make register_pending_environment return a PendingEnvironmentRegistration that is consumed by complete with either the stable URL or a terminal error.
• Represent pending WebSocket readiness as a shared one-shot result, awaited before entering the existing WebSocket connection and initialization path.
• Cache the shared terminal result so reconnects can resolve the same URL again.
• Remove the redundant URL copy and the public Environment::exec_server_url() getter. There are no in-tree production callers; test infrastructure now retains its own configured URL.
• Add exec-server integration coverage for successful completion, explicit failure, dropped or invalid registrations, and late completion after replacement.
• Update the deferred-executor core integration tests to register an environment without a URL, select it for the turn, and then exercise both late readiness and terminal provisioning failure.

This is an alternative implementation of #31988.

Testing

• just test -p codex-exec-server --test pending_environment
• just test -p codex-exec-server --lib (181 passed)
• just test -p codex-core --test all deferred_executor_updates_context_and_tools_after_startup
• just test -p codex-core --test all deferred_executor_wait_reports_startup_failure