Optimize reading physical time

[cmnrd]

After seeing this commit from @petervdonovan in the benchmark repository, I remembered that I implemented a simple optimization in the C++ runtime that completely avoids the overhead for reading physical time in the scheduler. The idea is to keep a cached value of the last read physical time. Let's call it last_pt. If the current logical time (let's call it lt) is smaller than last_pt, then there is no need to read the physical time again since we already know that whatever reading we will get, will be larger than lt. Only if lf > last_pt we read the current physical time and update last_pt.

Surely such an optimization could also be implemented in the C runtime. With this optimization, there is no measurable difference between benchmarks run with fast=true and without.