Investigate OOM for heavy query over datashards

[abyss7]

How to reproduce:

1. Initialize workload for TPCC: init, generate 10K DW.
2. Run workload with params: 2500 DW, 4 threads - on 3 nodes cluster, with 16 CPU User Pool.
3. Run query in parallel: https://paste.yandex-team.ru/16d434c2-9a3a-4333-a5fe-88c7f43e78da

In profile memory mode for soft_limit_percent=75 the following heap dump is available before OOM: https://paste.yandex-team.ru/49e09ccf-e642-4119-8085-e5c96eccb148