perf testing improvements

Some ideas that came out of the weekly storage meeting and a later discussion with @rmustacc

Collect dtrace output for the various parts and include that in the perf report. The idea being that whomever is looking at the report to see what changed, can dig into further details without having to re-run the test.

Get ptime -m results for not only the perf test itself, but the three downstairs as well.

Get dtrace like stats from each downstairs.
Gather downstairs timestamps at:

• IO received from upstairs (just off the wire)
• IO submitted to OS
• IO returned from OS
• IO ack'd back to upstairs (just before putting back on the wire)

Those will help us determine if there are issues in crucible downstairs, and if so, where.

For the graphs, a red line where any IO to the right is a problem.

For the output of the perf script, show 95 percentile and 99 percentile IOs instead of stdv/min/max.

Start by locating the performance test and report-generation entry points, then trace how the three downstairs components expose timing and DTrace-like data. The requested work covers ptime results, DTrace output, four downstairs timestamps, graph thresholds, and percentile reporting. It is done when the report contains these measurements and the graphs and percentile output reflect the requested criteria.