Introduce NumpyArray façade over py::array

[evertlammerts]

Add a thin wrapper class NumpyArray (src/duckdb_py/include/duckdb_python/ numpy/numpy_array.hpp) whose single data member is a py::array. This is now the only spot in the codebase that names py::array as the underlying numpy-array representation, so a future migration to nanobind's nb::ndarray is localized to this one header.

The façade exposes Data()/MutableData() (data buffer pointers), an Allocate() factory (dtype + count), a FromObject() factory, an explicit NumpyArray(py::array) constructor (a py::object argument implicitly converts via np.asarray semantics, matching prior behaviour), and GetArray() accessors for .attr(...) calls, iteration, resize, and handing the array back to Python. It is default-constructible, copyable, and movable.

Route every direct py::array use through the façade:

• numpy/raw_array_wrapper.{hpp,cpp}: member + Allocate/MutableData, resize via GetArray()
• pandas/pandas_bind.hpp (RegisteredArray) and pandas/column/ pandas_numpy_column.hpp: members + constructors take NumpyArray
• numpy/numpy_scan.cpp: scan helpers take NumpyArray&, .data() -> .Data()
• numpy/numpy_bind.cpp, pandas/bind.cpp: construct NumpyArray instead of py::array; dtype attrs via GetArray()
• numpy/array_wrapper.cpp (ToArray): move out / bool-check via GetArray()
• pyconnection.cpp, python_replacement_scan.cpp: py::castpy::array(...) ->
  wrap the object in NumpyArray and use GetArray()