- Epsilon is now understood as relative to expected value.
- Improve error messages for epsilon checks.

This is needed for examining the values of user counters (needed
for #348). It is also needed for checking the values of standard
benchmark results like items_processed or complexities (for example,
checking the standard deviation is needed for unit testing #357
as discussed in #362).

The tests are still missing a way to check actual validity of
numerical results; this will be done next. As they currently are,
the tests pass, but the problem detected with #378 is still
standing and the results with non-standard counters are wrong.

Fixes #378 coercion to double was causing counter to forget its flags, changed it so that its value is updated directly

The problem was that the call to Finish() the user counters was
lost in a big merge. If I had already written the tests for the
user counters, this would probably have been catched earlier.

* fix android compilation

* checking __GLIBCXX__ and __GLIBCPP__ macro in addition to __ANDROID__

* using vsnprintf instead of std::vsnprintf to compile on Android

* removed __GLIBCPP__ check on Android

* StringPrintF instead of std::to_string for Android

When using CPU time to determine the correct number of iterations the
library additionally checks if the benchmark has consumed 5x the minimum
required time according to the wall clock. This prevents benchmarks
with low CPU usage from running for much longer than actually intended.

However when a benchmark uses a manual timer this heuristic isn't helpful
and likely isn't correct since we don't know what the manual timer actually
measures.

This patch removes the above restriction when a benchmark specifies a manual
timer.