DOC: tweak np.round docstring to clarify floating-point error

[ahaldane]

Clarifies the docstring based on this discussion
https://mail.python.org/pipermail/numpy-discussion/2019-September/080034.html
https://mail.python.org/pipermail/numpy-discussion/2019-September/080044.html

To summarize, the behavior:

>>> np.round(16.055, 2), round(16.055, 2)  # equals 16.0549999999999997
(16.06, 16.05)

shows that floating-point error occurs for small values, not just large values like the np.round docstring says.

So this PR removes the red-herring comment about large values.

[mattip]

Thanks @ahaldane