ENH/PERF: allow mask to be optional in our masked ExtensionArrays

[jorisvandenbossche]

Our nullable, mask-based ExtensionArrays (currently integer and boolean, inheriting from MaskedArray) have a _data and _mask numpy arrays stored under the hood. SO we use a numpy boolean array as mask (8bit), also when there are no missing values.

One, relatively easy, memory + performance improvement could be achieved by allowing the mask to be None when there are no missing data. Since the mask data is completely internal to the Array implementations, this should be possible to do.

(to be checked how involved the ops code would become to handle this as optional)

[mzeitlin11]

@jorisvandenbossche @jbrockmendel @jreback planning to start looking into this, but want to check first that there's not a newly discussed preferred solution first.

[jbrockmendel]

not that im aware of

[jbrockmendel]

I think this suffers from the same invalidation problem as DTA/TDA.freq xref #31218. Consider:

arr = pd.array([1, 2, 3])
arr2 = arr[1:]
arr2[1] = pd.NA

If a mask isnt allocated until the value is set into arr2, then we need to make sure to allocate a mask for arr as well. Certainly possible, but not for the faint of heart.