Fix unbounded reconnect loop when all hosts fail during connect

[steobrien]

Connection's local retries counter is read in two places — error()'s options.host[retries + 1] guard and tryNext()'s options.host[retries] — and reset in the ready handler, but nothing ever increments it (closed() bumps options.shared.retries instead, and for connecting sockets it returns early before reaching that line anyway).

So with multiple hosts configured, a connect-phase failure always finds a "next host": error() swallows it, closed() reconnects, forever. The pending query never settles:

const sql = postgres({ host: 'localhost', port: 1 })
await sql`select 1` // rejects with ECONNREFUSED in ~10ms

const sql = postgres({ host: ['localhost', 'localhost'], port: [1, 1] })
await sql`select 1` // never settles

This change counts each failed attempt in closed() and stops reconnecting once the host list is exhausted. A socket error on the last host already rejects the pending query in error() with the original error (e.g. ECONNREFUSED), and a clean close with no hosts left now rejects with CONNECTION_CLOSED. The counter resets when the cycle ends, so the next query sweeps the full host list again; a successful connect resets it as before.

The added test mirrors the existing single-host ECONNREFUSED test — it times out on master and passes with the fix.