fix(client): harden supervised connection recovery edge cases

[enilsen16]

Follow-ups to #568 — a set of recovery edge cases found while reviewing that change after merge.

Connection (lib/grpc/client/connection.ex)

• Stale retry timers: schedule_retry deduped on a boolean but never cancelled a pending timer, so a transport death after a resolver-driven recovery could sit out the remainder of an old long backoff (up to ~2min) instead of redialing at the newly computed delay. Now stores the timer ref, cancels/reschedules, and adopt_established cancels any pending retry.
• Dead-channel adoption race: rebalance_after_reconcile adopted on connected != [] without the channel_alive? check the :retry_establish handler applies for exactly this race, so a channel that died with its :DOWN still queued could produce a false :connected (waiters released, telemetry emitted) until the :DOWN unwound it.
• Stale LB state for pickers: the LB state returned by lb_mod.update/2 was only kept in GenServer state; pick_channel reads :persistent_term, which was only re-put in establish/1. A behaviour-compliant LB that returns fresh state from update/2 (rather than mutating ETS in place) left pickers on the stale state indefinitely. Now republished when the state changes (no-op for the built-ins).
• Backoff ladder reset after flaps: adopt_established reset retry_attempt to 0 but not flaps, so an establish failure right after a flap-death restarted the dial ladder at the shortest delay — the backoff could shrink after a failure. handle_channel_down now seeds retry_attempt from the flap count.
• Silent resolver failures: non-{:ok, _} returns from resolver.update/2 were silently discarded (the removed handle_cast(:resolve_now) used to assert the match). Now logged, keeping the previous resolver state.

Stub (lib/grpc/stub.ex)

• Crash instead of UNAVAILABLE: fallback_channel's catch-all passed a payload like %{conn_pid: nil} — the documented return of Connection.disconnect/1 — straight to the adapter, crashing on a nil pid. Any payload carrying :conn_pid is now liveness-checked (nil/dead/remote → UNAVAILABLE); only payloads without the key pass through.
• Pid-less adapters treated as dead: resolve_channel used Map.get(payload, :conn_pid), conflating a missing key with nil, while channel_alive? treats the same shape as alive — so a healthy connection over an adapter that exposes no transport pid could never serve an RPC from the stub side. Now aligned with channel_alive?.
• Failure results bypassed connection config: unavailable_result ran the caller handle's (often empty) interceptor list with struct-default codec/compressor, while the healthy path uses the picked channel's config — so bare %Channel{ref: name} handles skipped the connection's interceptors on failures only. Now resolves config via Connection.get_channel/1 and honors the caller's codec/compressor opts.
• Wasted re-picks on non-rotating policies: the bounded re-pick loop burned all attempts on the identical dead channel under PickFirst (plus a Logger.warning per attempt). The loop now bails when the policy returns the same pid again.

Verification

mix test: 10 doctests, 360 tests, 0 failures (2 pre-existing skips). mix format --check-formatted clean on both files.

[sleipnir]

Thank @enilsen16