Fix and improve GRV batching

[Forums: How to scale FoundationDB reads](https://forums.foundationdb.org/t/how-to-scale-foundation-db-reads/1108/) has GRV benchmarks and results that show that proxy-side GRV batching seems broken, and that client-side batching is very conservative.

In a sense, this isn't a surprising result, as I'm not aware that we have a GRV benchmark. So we should
1. implement a GRV benchmark for latency at a give throughput, and throughput at a given latency for nightly performance tests. It's also worthwhile to add performance tests that are essentially GRV benchmarks as well: reads per second when one opens a new transaction for each read, and writes/s where one key is written per transaction.
2. Graph GRV latency vs GRV requests/s
3. Improve it

Start with the linked FoundationDB forum benchmarks and the repository's existing nightly performance tests. Define GRV latency at a target throughput, throughput at a target latency, and transaction-level read/write measurements before determining how to graph the results. Done means the benchmarks and graphs exist and GRV batching is improved based on their results.

Written by the indexing model from the issue text.