JSON reading: unified numbers

[Jolanrensen]

Fixes #557
Helps #961 by preventing Number columns from appearing in the first place, and by making it easier to create columns with unified number types.
Makes behavior also more consistent with other reading options, like CSV.

• Json can now read to Float. This was done for consistency with old behavior in the past (double parsing was checked before float parsing), but it can halve the memory usage of DataFrame in a lot of cases.
• Introduces unifyNumbers parameter to guessValueType and createColumnGuessingType.
• Json reading now uses unifyNumbers = true

[AndreiKingsley]

@Jolanrensen may be add unifyNumbers to readJson? I had an experience with APIs that returns JSON with different numeric types inside one column 😒 (and it was important).

[Jolanrensen]

@Jolanrensen may be add unifyNumbers to readJson? I had an experience with APIs that returns JSON with different numeric types inside one column 😒 (and it was important).

definitely, that sounds like a good idea :) I will make it true by default though. We want to generally avoid having Number columns unless it's absolutely necessary.

[zaleslaw]

I see that we are reducing Double -> Float what is possible, but why not Float -> Double as a most common type?

[Jolanrensen]

There's no number unification happening in this test at all; in fact, we're not reducing Double to Float on purpose, however because we're writing a Double "1.0" and "3.0" to JSON it can be read back from JSON as Float. If any of the numbers were too large to fit in a Float, or if the column contained both floats and integers, the result would have been Double.

[Jolanrensen]

@zaleslaw It's because of this change that we can now get floats out of json too instead of just doubles.