Edge case: duplicate step counts cause weird smoothing

If multiple points share a single step count, then smoothing can become very unrepresentative for a series.
Here's a simple artificial example:

And here's a more complex real world example of this happening:

This happens because the smoothing (a low pass filter) adjusts the smoothed series by a certain amount for every point that it processes. So if multiple points exist inside of a single step, then vertical lines will start to appear. Any data that comes after those vertical lines will be disproportionately influenced by those duplicate points, since they were "counted twice" (or 3 times, or 4, etc) in the smoothing filter.

I'll try to send in a PR to fix this when I have the time.

Start by locating the smoothing low-pass filter used by TensorBoard charts and reproduce the issue with multiple points sharing one step count. Trace how each point affects the smoothed series. Done means duplicate-step points no longer disproportionately influence later data or produce the reported vertical artifacts.