add delayed, bursty tracing control to drmemtrace

This is a feature request to add two parameters to drmemtrace (drcachesim's tracer):

First, -trace_delay N which counts N dynamic instruction executions and only then starts tracing.

Second, -detach_after_tracing N which traces N references (any metric: the current bb-start + memref for offline and instr+memref for online is fine) and then detaches. This will need #2644.

Together, these will allow making a "bursty trace" covering just a window of execution for an application launched under DR for which it's not easy to do a DR attach at the start point.