Keyword argument handling performance optimization.

Issue description

Keyword argument handling is slower than CPython.

CPython has an optimization that assumes the keyword name string objects are interned and performs pointer comparisons first instead https://github.com/python/cpython/blob/22424c02e51fab3b62cbe255d0b87d1b55b9a6c3/Python/ceval.c#L3975 .

Whereas Pybind constructs string objects and hashes them https://github.com/pybind/pybind11/blob/fb910ae92b4a489e9b31f9ec545bde9d09d6a40e/include/pybind11/pybind11.h#L518 .

What do you think about having a similar optimization in Pybind?

Reproducible example code

void matmul(py::handle a,
            py::handle b,
            py::handle transpose_a,
            py::handle transpose_b,
            py::handle adjoint_a,
            py::handle adjoint_b,
            py::handle a_is_sparse,
            py::handle b_is_sparse,
            py::handle name) {
}
 
PYBIND11_MODULE(test, m) {
 m.def("matmul", &matmul,
       py::arg("a"),
       py::arg("b"),
       py::arg("transpose_a") = false,
       py::arg("transpose_b") = false,
       py::arg("adjoint_a") = false,
       py::arg("adjoint_b") = false,
       py::arg("a_is_sparse") = false,
       py::arg("b_is_sparse") = false,
       py::arg("name") = "");
}

test.matmul(1, 2, transpose_a = True)

Start by reading the keyword-argument handling around include/pybind11/pybind11.h#L518 and compare it with the linked CPython ceval.c optimization. Run the reproducible matmul example to establish the current behavior and performance; done means the proposed interned-name or pointer-comparison optimization is implemented and its effect is verified.