Fixed multiple blocks finalization

[svyatonik]

Seems like it has been broken recently when atomic operations were introduced. So when GRANDPA finalizes several blocks at once they're enumerated here and when 2nd block is finalized, this check fails, because meta still contain previous best-finalized block.

I've also noticed that now we can't import block + finalize it in single operation, because commit first tries to finalize not-yet-inserted block. Also seems like there could be an issue with internal Backend' state because Backend' meta is updated first and then we could fail for example here => in-memory meta will be updated, while the database will not. So probably we need to do some refactoring (like introducing in-memory snapshots of Backend' state or separating finalize and import operations) to avoid issues like these in future. But I'm not sure if we need to support atomic import-and-finalize operations. If there's anyone with this knowledge, please let me know - I'll file an issue(s) for that ^^^.

[svyatonik]

Also probably it is better to replace in_mem backend in 'big' tests (that are using test_client) with a in-memory-database backend here, because in_mem backend is too specific && not used in production

[rphmeier]

Filed #1603 for improving the tests.

We will want to support atomic import-and-finalize ops, definitely