BUG: where gives incorrect results when upcasting (GH 9731)

[evanpw]

closes #9731

The main issue is when the destination and source arrays have different lengths, np.putmask doesn't behave like arr[mask] = values:

"Sets a.flat[n] = values[n] for each n where mask.flat[n]==True"

We have to use np.place instead. A secondary issue is that np.place doesn't automatically convert an integer to a datetime64 like np.putmask does (I created a numpy issue for this), so we need an additional check for that case. The rest of the commit is just cleaning up _maybe_upcast_putmask, which had some parameters that were never used, and a confusing docstring.

[jreback]

use lib.isscalar

[jreback]

@evanpw thanks!

this was subtle!

[jreback]

I assume you changed to use np.place because some OTHER tests were initially failing on your change. Or do we need some tests for the same operation but with dtype = M8[ns] or m8[ns]?

if so, can you do a pr with those?

[evanpw]

The tests that were failing after I changed np.putmask to np.place were the ones testing idxmax and idxmin, but the new tests in GH #9804 are more direct.