Fix rounding on float16 conversions

[emcastillo]

Closes #8366

When casting a Chainerx array in the native backend to float16 the float16 code differs subtly from the numpy one in the rounding of elements.

https://github.com/numpy/numpy/blob/1f8fb2ca31a7f41a23c60e4ffcb38b3e0cb201f7/numpy/core/src/npymath/halffloat.c#L339-L347

This caused the IRemainder test to yield 150 failures for 1 million runs.
After fixing the number went down to 0.

The fp64 to fp16 case has been similarly fixed to match the numpy logic as some of the calculations were different and it was not clear where did they come from.

[niboshi]

Also please write tests.

[emcastillo]

I added an initial try for testing it, it is for the float32 case,
If its ok I will extend it to take the float64 in account too.

Since this is a numpy related change I was doubting on how to do the testing.
The two alternatives that I though was to explicitly test if a number was not rounded, or
add a new python test which compares chainerx and numpy type castings. I went for the first one since the scope seems broader and the later its extensively tested in the routines.

[asi1024]

Jenkins, test this please.

[chainer-ci]

Jenkins CI test (for commit 51e1ec4, target branch master) succeeded!

[asi1024]

LGTM.

[chainer-ci]

Jenkins CI test (for commit 51e1ec4, target branch master) failed with status FAILURE.