kubernetes: wait for exec connection readiness

[jarqvi]

fixes #4040

The kubernetes driver dials a builder pod by running buildctl dial-stdio
through the exec API. ExecConn started StreamWithContext in a goroutine and
returned the connection immediately, before the SPDY stream was up. If the
upgrade failed (e.g. tls: internal error on a fresh node whose kubelet cert
isn't signed yet), Dial() had already reported success, so tryWithBackoff
never retried and the error surfaced later as a gRPC client-preface error.

Wait until the stream is established before returning. The executor reads stdin
only once its streams are created, so the first stdin read is used as the
readiness signal — no sleeps or probes. Transient setup failures are now
returned synchronously so tryWithBackoff can retry, and context cancellation
still unblocks the wait.

Added unit tests for newExecConn and tryWithBackoff using a fake executor.