More efficient support for linearizable reads

I believe the only way to achieve linearizable reads at the moment is for the app logic to create a phantom write to the kv to ensure that the primary replicates and reaches consensus on the associated txn. Without this, the read might be stale, even if forwarding is set to always and the client waits until the read txn ID is committed. This is because the primary who served the read may no longer be the current primary.

This is related to but somewhat separate from the discussion around wait for commit.

Here's some early ideas about how to make linearizable reads more efficient:

1. Allow txns with empty write sets - this probably the simplest approach. The app logic would still need to create a transaction for the read but at least with no write set, the transaction will have few conflicts and be smaller. It's a cleaner solution but I'm not sure how much is gained by this approach
2. Return the last sequence number +1 to the client for linearizable reads. This means that the client will not consider the read committed unless the primary commits another transaction (in the same term). The key issue here is that if the previous txn is a signature and no other write txns arrive then the client could be waiting indefinitely. This could be worked around by forcing another signature or using (1)
3. Implement PreVote or some type of leader stickiness such that if we trust clocks then the primary can serve linearizable reads without a round trip to a majority
4. Implement wait for commit and then delay responses to linearizable reads until after the next successful AppendEntries. This is closest to existing systems like etcd

_Aside: This could be an interesting to model in TLA+ using a separate high-level specification such as https://github.com/Azure/azure-cosmos-tla/blob/master/general-model/cosmos_client.tla_

Start by examining the existing handling of linearizable reads, forwarding, transaction IDs, and the related wait-for-commit discussion. Compare the four proposed approaches, including the referenced TLA+ model, and define which design should be implemented. Done means selecting and validating a concrete mechanism that avoids the current phantom-write requirement.

Written by the indexing model from the issue text.