common, osd, tools: Add histograms to performance counters

common, osd, tools: Add histograms to performance counters

This change adds new performance counter type: histogram.

Currently it does measure latency + request size (in bytes) of
an op similarly to l_osd_op_(r|w|rw)_lat and
l_osd_op_(r|w|rw)_(inb|outb). Histograms are 2-dimensional and
gather probes for both values at the same time. This allows
a more detailed analysis than what could be done with two
separate 1-D histograms. Still the memory footprint is negligible.

Since new data could break existing tools and the amount of data
dumped is rather large, two new admin socket commands are
introduced: perf histogram schema and perf histogram dump.
Invocation of original command should remain compatible.
There's also a new simple tool in src/tools/histogram_dump.py
which does a live dump of one histogram in local daemon.

Current configuration of histograms is hard-coded and should
cover all common cases. If it turns out to be useful, in the
future it could be loaded from the configuration.

Signed-off-by: Bartłomiej Święcki <bartlomiej.swiecki@corp.ovh.com>