Pass and receive numpy views in array-UDT UDFs - #589

A UDF over an array UDT wrote past the end of its output element. The cfunc
wrapper passed the raw element pointer and stored the return value with
`z_ptr[0] = ...`, so what landed in the buffer was Numba's NestedArray
descriptor (data pointer, shape, strides) rather than the element payload.
The descriptor is the wider of the two, so the store ran off the end of the
element into memory SuiteSparse owns. Measured against the code this
replaces, driving the wrapper's cfunc through ctypes over a guarded buffer:

    (6,)    np itemsize 48 | numba value type size 56 | overrun 8 bytes
    (2, 3)  np itemsize 48 | numba value type size 72 | overrun 24 bytes

    z[:6] = [100.0, 101.0, 102.0, 103.0, 104.0, 105.0]
    z[6:] = [5e-324, -1.0, -1.0, ...]       # -1.0 is the guard byte pattern
    indices written past the element: [6]

Both operands and the output are now bound with `numba.carray(ptr, shape)`
in the UDT's declared shape, so the UDF receives a numpy view it can index
(`x[i, j]`, `x.shape`) and the wrapper slice-assigns the result back with
`z[:] = ...`. Array-typed record leaves slice-assign for a second reason:
Numba's record-field setitem copies the destination's extent regardless of
the source's, so a short source was read past its end.

The return-type resolution has to move with the wrapper. Once the UDF
receives a view, a UDF that builds its result (`x + y`) instead of returning
an operand types as a plain Numba `Array`, which `lookup_dtype` does not
recognize; `_resolve_udt_return_type` gains an `Array` branch matching it
back to an input array UDT by base element type and rank. Splitting the two
apart would leave array-UDT UDFs that build a result broken outright.

An `Array` type carries `ndim` but not its extents, so a return whose shape
does not fit the UDT is not a type error. It is now memory-safe, but still
silent: the slice-assign's ValueError is raised inside a cfunc, which Numba
prints and swallows, leaving the element as SuiteSparse found it.