Support Eigen::Blocks?

[adriendelsalle]

Hi!

First, thanks for this project :)

Description

It looks like the current implementation doesn't support the dense block API of Eigen.
It would be awesome to handle this!

Tests

Adding the following tests:
test_eigen.cpp

    nb::class_<ClassWithEigenMember>(m, "ClassWithEigenMember")
        .def(nb::init<>())
        .def_rw("member", &ClassWithEigenMember::member)
        .def("get_block", [](ClassWithEigenMember& c){ return c.member(Eigen::seq(0, 1), Eigen::seq(0, 1)); }); 

test_eigen.py

@needs_numpy_and_eigen
def test12_block():
    c = t.ClassWithEigenMember()
    for mat in (np.ones((2, 2)), np.ones((3, 3))):
        c.member = mat
        block = c.get_block()
        ref = np.ones((2, 2))
        assert block.size == 4
        assert block.shape == (2, 2)
        assert np.all(block == ref)

        block[0, 0] = 10.
        assert c.member[0, 0] == 10.

leads a TypeError:

>           block = c.get_block()
E           TypeError: Unable to convert function return value to a Python type! The signature was
E               get_block(self) -> Eigen::Block<Eigen::Matrix<double, -1, -1, 0, -1, -1>, -1, -1, false>

Discussion

As far as I understand there are type_casters for both Eigen::Map<T>

nanobind/include/nanobind/eigen/dense.h

Lines 175 to 177 in 5e6b210

	/// Caster for Eigen::Map<T>
	template <typename T, int Options, typename StrideType>
	struct type_caster<Eigen::Map<T, Options, StrideType>, enable_if_t<is_eigen_plain_v<T>>> {

and Eigen::Ref<T>

nanobind/include/nanobind/eigen/dense.h

Lines 236 to 238 in 5e6b210

	/// Caster for Eigen::Ref<T>
	template <typename T, int Options, typename StrideType>
	struct type_caster<Eigen::Ref<T, Options, StrideType>, enable_if_t<is_eigen_plain_v<T>>> {

not covering the Eigen::Block<T> one.

Naively Eigen::Map<T> could/should cover Eigen::Block<T> since both are derived from Eigen::MapBase<T>? I can propose a PR on that!
There is also this other discussion and associated PR on Eigen topic, but it looks like the scope is much larger.
Let me know!