Implement `ArkoudaCategoricalArray.__setitem__` (pandas-aligned)

[ajpotts]

Summary

Implement ArkoudaCategoricalArray.__setitem__ to support
pandas-compatible item assignment into Arkouda-backed categorical
ExtensionArrays.

This is required for common pandas workflows such as:

• Series.loc[...] = ... / Series.iloc[...] = ...
• boolean mask assignment
• where/mask
• fillna and other in-place manager paths
• categorical value replacement without dtype loss

Currently, assignment into Arkouda categorical arrays is missing or
inconsistent, leading to TypeError/NotImplementedError or pandas
fallback behavior (often converting to object/NumPy).

---

Background / Why

pandas Categorical supports item assignment with strict rules:

• Assigned values must be existing categories or missing
• New categories are not implicitly added (unless user explicitly
  adds them via add_categories or similar higher-level API)
• Missing values are supported and propagate through codes/mask
• Assignment must preserve dtype
  (CategoricalDtype(categories=..., ordered=...))

For Arkouda-backed categoricals, we want identical semantics while
keeping operations server-side where possible.

---

Requirements / Expected pandas Semantics

Given categories ["a", "b"]:

1. Assign existing category:
   • cat[0] = "b" is allowed
2. Assign missing:
   • cat[0] = None / pd.NA is allowed and marks entry missing
3. Assign value not in categories:
   • cat[0] = "c" should raise (typically
     TypeError/ValueError depending on path)
   • pandas message often indicates: "Cannot setitem on a Categorical
     with a new category..."
4. Assignment via indexers should work:
   • int, slice, boolean mask, integer array indexer
5. Broadcasting rules:
   • scalar value broadcasts to all targeted positions
   • array-like values must match number of targeted positions

---

Scope

In Scope

• Implement ArkoudaCategoricalArray.__setitem__(key, value)
• Support keys:
  • int position
  • slice
  • boolean mask (same length)
  • integer indexer (array-like positions)
• Support values:
  • scalar category label
  • scalar missing (None, pd.NA, possibly np.nan)
  • array-like of labels/missing matching target selection length
  • another ArkoudaCategoricalArray (assignment by position)
• Enforce "no new categories" rule
• Preserve:
  • categories
  • ordered flag
  • dtype and internal representation (codes + categories + missing
    marker/mask)
• Add unit tests

Out of Scope

• Adding categories automatically during setitem
• Implementing add_categories / remove_categories (if not already
  present)
• 2D assignment (categorical EA is 1D)
• Alignment by Index labels (handled by pandas, not EA)