StreamLookup is going into a spinning frenzy on slow resolve after split

[snaury]

Example test and log where datashard splits and resolve is blocked when datashard replies with NOT_FOUND: https://paste.yandex-team.ru/e95fe27d-6d14-46c4-beb3-ecd3d4f696c9

This apparently causes compute actor to go into a spinning frenzy, repeatedly asking for more rows, and the test cannot simulate sleep and unblock the resolve later, because actor system never stops.