GRANDPA pre-vote selection algorithm

[AlistairStewart]

The spec for prevoting is inconsistent with Substrate and the GRANDPA paper.

Best-PreVote-Candidate returns either the primaries block or the ghost of the prevotes apparently from this round r.

What the validator should prevote for is for the last audited block (or 3 blocks from the end in single-chain Substrate) in the longest chain (or chain with the most primary blocks for BABE) including a block B defined as follows

If the primaries' block is between the last finalused block L and the ghost of the prevotes from round r-1, then that is B

else use the ghost of the prevotes from round r-1.

We should probably check whether Substrate works like this.

[AlistairStewart]

I think 3 things need fixing:

1. In Best-PreVote-Candidate, GRANDPA-GHOST should be called on round r-1 and the resulting variable renamed to include r-1.
2. At the end of Best-Prevote candiate, we should call Best-Chain(N)
3. Best-Chain should defined somewhere.

[AlistairStewart]

We still need a best chain algorithm somewhere.