[gateway] Share dial failures with current waiters

### Description

PR #4078 intentionally keeps connection management simple: a failed dial is not cached, so waiting and subsequent requests retry after the previous attempt finishes.

As a follow-up, evaluate sharing one completed dial failure with requests that were already waiting for that attempt, without caching the failure for future requests. Later requests should still retry immediately after the failed attempt completes.

Start by reviewing PR #4078 and the gateway connection-management path it changed. Trace how waiting requests observe a failed dial, then define completion as sharing that failure only with current waiters while later requests retry immediately without cached failure state.

Written by the indexing model from the issue text.