Do not lose precision for durations specified without fractions
that can be represented by an int64 such as 1<<53+1 nanoseconds.
Previously there was some precision lost in floating point conversion.

Handle overflow for durations above 1<<63-1 nanoseconds but not earlier.

Add tests to cover the above cases.

Change-Id: I4bcda93cee1673e501ecb6a9eef3914ee29aecd2
Reviewed-on: https://go-review.googlesource.com/2461Reviewed-by: Russ Cox <rsc@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>