Pybind11 dtype caster does not accept strings or type objects

Issue description

If you bind a function where one of the arguments is a pybind11::dtype, the function does not automatically cast str objects or type objects to dtype. In contrast, all Numpy functions accept str or type arguments for their dtype: e.g. np.zeros(10, dtype=float) or np.zeros(10, dtype="float32") are both valid.

Reproducible example code

When I write:

#include <pybind11/pybind11.h>
#include <pybind11/numpy.h>

PYBIND11_MODULE(mymodule, m) {
  m.def("test_dtype_cast", [](pybind11::dtype t) {
    pybind11::print(t);
  });
}

and call:

import numpy as np
import mymodule as m
m.test_dtype_cast("float32")

I get the error:

TypeError: test_dtype_cast(): incompatible function arguments. The following argument types are supported:
    1. (arg0: dtype) -> None

Invoked with: 'float32'

Similarly, when I call

import numpy as np
import mymodule as m
m.test_dtype_cast(np.float32)

I get an error:

TypeError: test_dtype_cast(): incompatible function arguments. The following argument types are supported:
    1. (arg0: dtype) -> None

Invoked with: <class 'numpy.float32'>

Both these cases should get implicitly casted to pybind11::dtype.

Start with the reproducible PYBIND11_MODULE example and inspect how pybind11::dtype arguments are converted. Compare the shown string and numpy.float32 calls with NumPy's accepted dtype inputs; done means both forms are accepted when passed to test_dtype_cast, with coverage for the reported cases.