[tracing] Make TRACE_DURATION use trace duration complete events

... instead of a start and end events. Saving ~50% in buffer and JSON
sizes.

There is no noticeable change in performance for NTRACE, tracing
disabled, category disabled and tracing enabled.

The compiled code size is less than or nearly equal the old version for 4
or less arguments but it is bigger for 8 arguments and more, issue
tracked in PT-87.

TEST: fx shell /system/test/sys/trace-test
TEST: fx shell trace-benchmark

PT-72 #done

Change-Id: I28757309cf62f6a3a84a9263538ff4dc503bae1f