[Filestore] Hardlink creation in sharded filesystems is not atomic

[qkrorlqr]

Hardlink creation in sharded filesystems is done via this actor:

nbs/cloud/filestore/libs/storage/service/service_actor_createnode.cpp

Line 25 in 5329d25

class TLinkActor final: public TActorBootstrapped<TLinkActor>

It first creates a hardlink in the shard which is responsible for the linked node and then creates a link in the main tablet. This whole operation is not atomic - the first step may succeed and the second may fail after that. We need to reimplement it in a way similar to regular node creation - first step should add a request to OpLog in the same transaction with the other logic of this step and the second step should be retried until it either succeeds or fails with a non-retriable error. The difference with regular node creation is that here the order is reversed - we first create a node in the shard and, second, we make modifications to the main filesystem. Those modifications may fail so extra logic may be needed - i.e. rollback of the first step in shard.