Improve the model of CPU work required for the proxy to commit

Currently the model is that a proxy cannot commit more than 200 MB/s, however the actual work is heavily dependent on the ratio between conflict ranges and mutations, and also the number of mutations rather than just the number of bytes in the mutations.

A more accurate model will help the proxies avoid fetching versions from the master before they have the CPU available to process the batch, which will improve throughput and lower tail latencies.

Related to https://github.com/apple/foundationdb/pull/2589