PERF: json should process data column-by-column (and not use .values)

[jreback]

xref #9027

So to_json first converts all the a DataFrame by using .values. This converts everything to object dtype (if its a mixed frame) and is pretty expensive perf wise.

If you do this column by column you get excellent perf (and tiny bit more complexity in the code).

cc @Komnomnomnom
@cpcloud

[Komnomnomnom]

Thanks @cpcloud for #9028

I should have some time to play with this this week. Will there be any perf impact going column by column with a non-mixed frame?

[jreback]

@Komnomnomnom

you should be able to go column by column for any kind of frame (and perf is essentially the same) except if its very large.

So the 'ideal' way to do this is actually

for dtype, block in df.blocks():
    # this is a single dtyped frame

So you get costless single-dtyped frames

[jreback]

closed by #9130