Use sampling for faster measuring of CPU time

Calling SystemClock.currentThreadTimeMicro() is expensive (order of
1 microsecond).

To mitigate that, we use sampling by measuring every 10th call for the
given class/transaction combination and extrapolate results.

Sampling is not used when --enable-detailed-tracking option is set,

For more compact output, only 90th percentile is printed, unless -a
option is set. Example output:

Per-UID Summary (cpu_time, % of total cpu_time, call_count, exception_count, package/uid):
    13095032  43%    44663   0 com.package1/u0a35
     8872156  29%     5831   0 com.package2/u0a75
     2379684   8%     1549   0 com.package3/u0a23

Test: manual dumpsys binder_calls_stats
Test: BinderCallsStatsTest + BinderCallsStatsPerfTest
Bug: 75318418
Change-Id: I535e796c17e75d70970e227411cdf0e1a4afd81c