Fix two races in controller.rs - #6786

The backpressure thread recorded an endpoint id in `running_endpoints` before
checking for that endpoint's reader, so an endpoint that was registered but not
yet opened lost its only `extend()`: `insert` returned false on every later
scan, leaving the endpoint paused forever. Skip a reader-less endpoint instead,
and the scan that follows `add_input_endpoint`'s unpark starts it.

`add_input_endpoint` registers an endpoint with `reader: None` and installs the
reader only after `open` returns, a window open since d24ba40. Only endpoints
added to a running pipeline can land in that window, because the connectors
named in the pipeline configuration are all created before the backpressure
thread starts. Every HTTP `/ingress` push creates such an endpoint, and the
endpoint blocks in `complete_request` until `extend` arrives, so a lost command
hangs the request until the client gives up. That is how
`server::test_http::test_concurrent_bootstrap` failed in CI, where the client
timed out after 120 seconds, and the same window can hang an ingress request in
production.

Signed-off-by: Leonid Ryzhyk <ryzhyk@gmail.com>