Fix percentile calculations

Adjust max latency in running binderThroughputTest
to improve percentile accuracy. We used to use a fixed
max latency which is devided into time buckets to obtain
the round trip latency percentile. The actual latency,
however, can be much smaller than the max latency, which
results in measurements from all iterations ending up in
the same bucket and leads to inaccurate percentile
measurement. This is resolved by either performing a
training round (option -t), which sets the max latency to
two times the worst latency during training, or by providing
a max latency (option -m) in millisecond by the caller.
Use --help option to see usage.

Bug: 62660944
Change-Id: I23899f936a9eca8ce6935fb9d554a6465bf2c797
Merged-In: Iaa23837eb24b90012ae20c35f20caae33986b35a
Fixes: 62660944
Test: Running "binderThroughputTest -p -w 2  -i 10000 -s 0" on marlin-3.18 used to show the same percentile values for 50%, 90%, 95%, and 99%. Now with -t flag,percentile values are different.
Signed-off-by: Sherry Yang <sherryy@google.com>