Implement rounding mode for Number#round

[straight-shoota]

This patch adds Number#round(RoundingMode) which enables to chose rounding behaviour at runtime.

Supported rounding modes (per #10228 (comment)):

• TO_ZERO
• TO_POSITIVE
• TO_NEGATIVE
• TIES_EVEN
• TIES_AWAY

Implementations were already available as dedicated methods for all modes except TIES_EVEN which is added as #round_even.

#round_away is the explicitly named implementation of TIES_AWAY mode previsouly just named #round.

#round without arguments now delegates to #round_away and the default rounding mode is TIES_AWAY for keeping backward compatibility.

Resolves #10228

There's already #7126 waiting with the implementation for big numbers (it will require some updates, though).

[bcardiff]

Looks good. I would feel more comfortable with ceiling/floor. I think they are more familiar names. But I guess that sticking with IEEE-754 is fine.

[straight-shoota]

The failing spec depends on #10413.

[bcardiff]

I guess you meant #10420, right?

[Sija]

I'd move it to the beginning of the file for a better discoverability.

[straight-shoota]

What difference does it make? I think it should stay in proximity to the methods that use it.

[straight-shoota]

We could even consider extracting all features related to rounding into a separate file.

[Sija]

We could even consider extracting all features related to rounding into a separate file.

Sounds even better.

[straight-shoota]

Let's leave that for a follow-up.

[HertzDevil]

This breaks BigFloat:

require "big"

BigFloat.new("1.23").round # Error: undefined local variable or method 'round_away' for BigFloat

I'm not sure if anything can be done about this, because GMP doesn't implement TIES_AWAY and TIES_EVEN. MPFR does though (plus a whole bunch of other functions suitable for Math). We most certainly don't want to make the default something else only for BigFloats.

[straight-shoota]

This is similar to #10599 and should probably be fixed by #7126.

[HertzDevil]

It will not be fixed because GMP straight up doesn't define these rounding modes, and we cannot implement them for BigFloat on top of either BigDecimal or BigInt.

BigDecimal is, to my knowledge, not a drop-in replacement for BigFloat.

[straight-shoota]

Oh, well. Then there is probably nothing we can do? The error is probably good then. You can't use BigFloat#round without rounding mode argument because the default mode is not implemented. We should make sure that using other rounding modes explicitly works, though.