Add `pd.NA` Handling to `ArkoudaArray._arith_method`

[ajpotts]

Summary

Extend ArkoudaArray._arith_method to explicitly handle pd.NA
operands.

Currently, when an arithmetic operation involves pd.NA (e.g.,
array + pd.NA), _arith_method returns NotImplemented. pandas then
falls back to a slower path, often materializing the Arkouda-backed
array as a NumPy array, leading to:

• Loss of dtype preservation
• Performance degradation
• Potential semantic inconsistencies

This ticket ensures proper missing-value propagation for arithmetic
operations while keeping computation server-side.

---

Background / Problem

pandas dispatches arithmetic operations on ExtensionArrays through
_arith_method.

Example:

s = pd.Series([1, 2, 3], dtype="ak")
s + pd.NA

Current behavior: - _arith_method sees pd.NA - Returns
NotImplemented - pandas falls back to NumPy/object logic - Arkouda
array may be converted to NumPy - Performance and dtype guarantees are
lost

Desired behavior: - Arithmetic involving pd.NA should remain within
Arkouda-backed logic - Missing values should propagate according to
pandas semantics

---

Expected pandas Semantics

Arithmetic with pd.NA generally results in missing values propagating:

• x + pd.NA → all results missing
• pd.NA + x → all results missing
• array_with_missing + pd.NA → all missing
• array + scalar where array contains missing → missing preserved
  elementwise

Behavior should match pandas nullable dtypes.

---

Scope

In Scope

• Modify _arith_method to detect pd.NA
• Implement correct missing-value propagation
• Preserve Arkouda dtype and mask representation
• Avoid NumPy fallback
• Add regression tests

Out of Scope

• Redesigning arithmetic dispatch
• Changing existing non-NA arithmetic logic
• Supporting arbitrary third-party NA-like objects