As the benchmark policy and the interval statistics start at
nearly exactly the same time this could cause the interval statistics
to report one interval too much or not.

This change slightly delays to policy start and suppresses printing
of the last incomplete interval. That way the first interval captures
slightly too little data (but that doesn't matter as it's normally cut
off anyways) and the last incomplete interval only captures data
for the fraction of a second.

This ensures that a local configuration test will trigger these checks.

This defers the creation of request paths until a
client is actually instantiated. Previously this initialization
happened when REFITConfig loaded the client class, which
slows down the configuration check.

When passing '--check' as first argument, then an experiment
will run a basic configuration check.