Unique index contraint violation at update can lead to an inconsistent state

[Totktonada]

We perform delete, then insert in accessor_shard.update_tuple, so the conflict will occur on the insert operation and the delete will not be rolled back.

@opomuc proposed to make insert and then delete.

[opomuc]

After brainstorming found only one case, when proposed behavior might not lead to expected result. When updating parts of a unique index (including, but not limited to, shard_key) to themselves (same value) a user might expect the same behavior as Tarantool's space:update() but instead will recieve an error. Cross-shard update in this case also seems concerning

[Totktonada]

@opomuc This case can (and must) be avoided using check that we insert an updated tuple into a storage instance that is different from one from which we’ll delete the tuple then. I’ll leave detailed comments for update_tuple function and will give the PR to your review.