Right now we do unit conversions between PerfCounter measurements
and nanoseconds for every add/sub we do between Durations and Instants
on Windows machines. This leads to goofy behavior, like this snippet
failing:

```
let now = Instant::now();
let offset = Duration::from_millis(5);
assert_eq!((now + offset) - now, (now - now) + offset);
```

with precision problems like this:

```
thread 'main' panicked at 'assertion failed: `(left == right)`
  left: `4.999914ms`,
 right: `5ms`', src\main.rs:6:5
```

To fix it, this changeset does the unit conversion once, when we
measure the clock, and all the subsequent math in u64 nanoseconds.

It also adds an exact associativity test to the `sys/time.rs`
test suite to make sure we don't regress on this in the future.

Per review comments, this commit switches out the backing
type for Instant on windows to a Duration. Tests all pass,
and the code's a lot simpler (plus it should be portable now,
with the exception of the QueryPerformanceWhatever functions).