One poller owns one concurrency slot: it long-polls /runtime/invocation/next,
runs the handler and posts the result, then polls again.
Standard Lambda runs exactly one poller. On Lambda Managed Instances
Mayfly.Supervisor starts AWS_LAMBDA_MAX_CONCURRENCY of them; each poller
is an independent process, so invocations never share state except the
handler state returned from init/1.
Poll failures are retried with exponential backoff (100 ms to 5 s) and
emit [:mayfly, :poll, :error]. Non-recoverable Runtime API answers
(HTTP 500 "container error") stop the poller, which makes the supervisor
shut the VM down as the Runtime API contract requires.