TypedPython: numpy array parameters (types, warnings and division follow numpy scalars) #220

Split from #149. array.array buffers compile zero-copy, but numpy arrays cannot simply reuse that path without breaking CPython identity. Verified with numpy 2.5.4 on 3.15t:

• In the interpreter, element arithmetic produces numpy scalars (np.float64, np.int64, np.bool); s += v over the elements returns np.float64. Compiled code would return float.
• numpy scalar ops emit RuntimeWarnings (overflow, invalid value), and division by zero warns instead of raising.

Design needed:

• taint tracking of values derived from numpy elements;
• boxing returned or escaping tainted values as the matching numpy scalar type;
• reproducing numpy's floating-point warnings (FP exception flags turned into the same RuntimeWarning per operation site);
• numpy's division semantics on tainted values.

Done when: compiled numpy functions match the interpreter in results, result types and emitted warnings (checked with warnings.catch_warnings(record=True)), and the nbody-style 7-array functions compile.