Caller-supervised shared-channel cache for Arcadic.Transport.Grpc.

A gRPC channel is HTTP/2-multiplexed — many concurrent streams share ONE connection — so this is a shared channel cache (one long-lived channel per {host, port, tls?, trust-selection} endpoint reused across ALL calls with the SAME trust config), NOT an exclusive checkout pool that would serialize streams. It opt-in: add

children = [{Arcadic.Transport.Grpc.ChannelPool, []}]

to your supervision tree. When it is running, the transport reuses its channel; when it is absent, the transport falls back to a fresh per-call connect (no behavior change). Tenant-blind — keyed on the endpoint + trust selection only, carrying no database/scope. Rationale for the process: a live gRPC channel is a socket/connection process (adapter_payload.conn_pid), not an immutable value copied into a pure-data Conn — sharing one requires a process, which is why the pool is opt-in and caller-supervised rather than transport-internal.

The CONNECT itself runs in the CALLER's process, never inside the GenServer: a connect can block for the adapter's full connect timeout (grpc 1.x default 15s, ×2 with the transport's transient retry), and serializing that inside the pool would queue every OTHER endpoint's checkout behind one dead endpoint. The GenServer only makes two O(1) decisions — a cache lookup, and a first-wins register when two concurrent first-connects race (the loser's channel serves its own call and is then best-effort disconnected). A dead channel (its adapter connection process gone) is transparently replaced on the next checkout.