The variance of the returned timing is reduced.
Now the median of multiple measures is returned in stead of the mean, so that outliers have less impact on the result.
The function also runs for at least 3 consecutive times (or 50ms), giving the possibility to measure the overlapping between CPU and GPU. This can result in lower, although more correct readings, especially for fast functions.

Description

When measuring an empty function {}, the returned timing was 1/700 lower (1.2ps vs 800ps). This is, because the number of cycles is automatically increased to obtain 50ms for each measure.
For slow functions, at least 3 cycles are used for the measure.
In total 0.5s is the typical total time for the function to return.

The variance has now become small enough, so that also faster functions can be compared on speed progress.
A comparison, between old and new, on af::constant() can be found in timeit.xlsx

• Is this a new feature or a bug fix? Improvement
• Why these changes are necessary. The variance on small arrays is too high for correct comparison.
• Potential impact on specific hardware, software or back-ends. None
• New functions and their functionality. No change to API, faster is most cases.
• Can this PR be back-ported to older versions? Yes
• Future changes not implemented in this PR.

Changes to Users

The returned time will be of higher quality.

Checklist

• Rebased on latest master
• Code compiles
• Tests pass
• [-] Functions added to unified API
• [-] Functions documented