load_gen improvement w/ more statistics

[emaxerrno]

Currently the RPC project comes with built-in load generator and statistics of latency only.

It also has average bytes per second per core and other minor ones.

We need new framework for extending the statistics

Facebook's Threadmill https://github.com/facebook/treadmill/blob/master/ContinuousStatistic.cpp

has an interesting design we should study and see what would could be ported in a way that is:

1. Seastar friendly (no locks, atomics, synchronization)
2. Does not add too much overhead to the measured system

Our histograms are very friendly as we preallocate 158KB per core - for all latency metrics, so updating them is very cheap

Future additional improvements also from Iago: twitter's load tester is different load distributions.

exponential, sustained, spiky, etc

with a WARM-up step

Currently we measure latencies of ALL requests. This is problematic because the first connection is very costly - usualy - since we do lazy initialization of most things, including cache fetching etc.

[emaxerrno]

Duplicate of #257