Add a section to the docs that compares cuDF with Pandas

[shwina]

Adds a section to the docs that calls out the similarities and differences from Pandas at a high level.

This is inspired by CuPy's page documenting the differences from NumPy.

[brandon-b-miller]

I think this is a great place to call out the subtle differences in null handling logic we have vs pandas. Most of it can be dug up from the source code here but a good summary might be something like this (I think this is all of them?)

Nulls in cuDF behave differently from pandas in several edge cases. In cuDF, the rule is that nulls always propagate, whereas in pandas they may not if the mathematical result can be inferred without knowing the missing value: 
- `NA ** 0 == 1`
- `1 ** NA == 1`
- `NA | True == True`
- `True or NA == True`
- `False and NA == False`

Maybe a table or something might be better than this.

[bdice]

All these cases are also described in the docs (as a cross-reference with the source code linked above):

• https://pandas.pydata.org/docs/user_guide/missing_data.html#propagation-in-arithmetic-and-comparison-operations
• https://pandas.pydata.org/docs/user_guide/missing_data.html#logical-operations

I find it a little concerning that we differ in this way because it means that cuDF cannot be consistent in its behaviors between scalars and columns. It should be specifically noted that scalar operations act like Pandas (because we use the same magic NA singleton object), and column operations always propagate NA.

>>> import cudf
>>> cudf.NA ** 0
1
>>> cudf.Scalar(cudf.NA, dtype=float) ** 0
Scalar(1.0, dtype=float64)
>>> cudf.Series([cudf.NA], dtype=float) ** 0
0    <NA>
dtype: float64

[shwina]

Yeah the difference in column vs scalar behaviour is problematic. I think @brandon-b-miller has thought a lot about this, where maybe we should take this discussion offline and come back and raise a separate issue if needed.

For this PR, I'll hold off on adding any further information about null behaviour.

[vyasr]

I would recommend thoroughly reading the discussion on pandas-dev/pandas#29997 before we relitigate any of that discussion.

[shwina]

Based on our discussions offline, I'm going to hold off on documenting the exceptional cases here. I think our priority should be to first align the behavior of nulls in all three of the following cases:

• Scalar operations involving NA
• Column operations involving NA
• Operations in UDFs involving NA

We can choose to always return NA in all three cases, or make an exception for ** in all three cases, but we must be consistent. That done, we can come back here to document the difference from Pandas - if any.

[bdice]

LGTM with some small edits. Thanks @shwina!

[brandon-b-miller]

marking my approval for now since my only concern was around the null behavior.

[shwina]

@gpucibot merge