Cudf: [FEA] Introduce `DecimalDtype` to cuDF

[shwina]

As a first step towards supporting decimal fixed-point arithmetic, we need to introduce a new DecimalDtype dtype.

Requirements:

• Similar to other cuDF dtypes, should inherit from ExtensionDtype. This requires us to define the abstract methods name and type. We could probably use 'decimal' and decimal.Decimal respectively.

• Similar to PyArrow's Decimal128Type, should be constructed from a precision and a scale attribute:

  >>> import cudf
  >>> dt = cudf.DecimalDtype(precision=4, scale=2)
  

• Should expose .from_arrow() and .to_arrow() methods that convert from and to PyArrow Decimal128Type.

[harrism]

Should it have the explicit storage size in the name like PyArrow? libcudf has 32-bit and 64-bit now, with plans to add 128-bit in the future.

[harrism]

@kkraus14 should this aim for 0.18?

[kkraus14]

Should it have the explicit storage size in the name like PyArrow? libcudf has 32-bit and 64-bit now, with plans to add 128-bit in the future.

In the call we had from the Python side I think we decided to only have a single DecimalDtype for now hooked up to DECIMAL64.

Then as a follow up we'll explore automatically handling type promotion / demotion based on the calculated precision value from operations.

If that exploration doesn't pan out we'll split the types into Decimal64Dtype and Decimal32Dtype.

@kkraus14 should this aim for 0.18?

No I think we can safely do this in 0.17.