Binder calls stats - random sampling.

Change the sampling method to be random. The previous mechanism recorded
one call for each key (uid/API name) and every X calls for each key:
- This is biased and will make it to interpret the data from
westworld.  It was especially unfair for apps using many different APIs
since the first call to each API was always recorded.
- It uses more memory since we will keep track of all the long tail

Simplify/unify the way we keep track of sampled calls.

Do not estimate the CPU usage of non-recorded calls
    long samplesCount = cs.callCount / mPeriodicSamplingInterval + 1;
    duration = cs.cpuTimeMicros / samplesCount;
It biases the results, let's use an example with 3 calls: 1ms, 3ms, 5ms
with an sampling interval of 2. With the previous algorithm we would get an
average per call of (1+1+3+2+5)/5=2.4ms. With the new one (1+3+5)/3=3ms.

Test: unit tests
Change-Id: I1dd7eb3c6c631b86a53485bccbfa397882cccc92