fix(api): treat col == None or col == ibis.NA as col.isnull()

[jcrist]

Previously we coerced co == None to IdenticalTo(col, None), but wouldn't do so for ibis.NA (resulting in that compiling as Equals(col, None) (which would result in incorrect SQL). This PR fixes this eager branch to handle any null literal, and compiles it the same as isnull/notnull (usually col IS NULL) rather than the more verbose col IS NOT DISTINCT FROM NULL that using IdenticalTo would result in.

Other options for this fix would be to:

• Always return Equals and use a rewrite rule to rewrite these cases. I opted against this since we'd have to have the same detection logic later, and this rewrite rule would be applied for every backend. Honestly having Equals.__new__/NotEquals.__new__ optionally return IsNull/NotNull would be cleaner IMO, but that's harder to do with our operation infrastructure.
• Always return Equals, and branch when compiling to generate the proper SQL there. This is a bit harder to do, and also would need to be implemented for dataframe backends so col == None would still use the proper isnull call.

[cpcloud]

How would someone compute a NULL-preserving equality with this change? It seems like it would no longer be possible to do that.

For the semantics you've defined here, you can already do this with a.identical_to(NA).

[cpcloud]

Oh, it looks like we're doing that already (as you note in the description), nevermind!

[cpcloud]

Seems like the reversed operand case doesn't work. Looking into it.