add support of dynamically sized Eigen types

[WanZhiQiu-ac]

I use b[ix++] = '['; because I saw something like

if (const auto k = ix + write_padding_bytes; k > b.size()) [[unlikely]] {
   b.resize(2 * k);
}

so this might be safe. Not so sure.

A quick test:

#include <Eigen/Core>
#include <iostream>

#include "glaze/ext/eigen.hpp"
#include "glaze/glaze.hpp"

class Example
{
   int a{1}, b{2}, c{3}, d{4}, e{5}, f{6}; // ... and many many more

   friend struct glz::meta<Example>;

  public:
   Eigen::MatrixXd mat;
   Eigen::VectorXd vec;
};

template <>
struct glz::meta<Example>
{
   using T = Example;
   static constexpr auto value = object(&T::a, &T::b, &T::c, &T::d, &T::e, &T::f, &T::mat, &T::vec);
};

int main()
{
   Example obj1{}, obj2{};
   obj1.mat.resize(2, 3);
   for (int j = 0; j < 2 * 3; ++j) {
      obj1.mat(j) = j + 1.01;
   }
   obj1.vec.resize(3);
   obj1.vec << 1, 2, 3.33333;

   std::string buffer = glz::write_json(obj1).value();
   std::cout << buffer << '\n';

   assert(obj2.mat.size() == 0);
   assert(obj2.vec.size() == 0);

   auto ec = glz::read_json(obj2, buffer); // populates s from buffer
   if (ec) {
      std::cout << "error" << std::endl;
   }
   assert(obj2.mat.rows() == 2);
   assert(obj2.mat.cols() == 3);
   assert(obj2.vec.size() == 3);
   assert(obj1.mat == obj2.mat);
   assert(obj1.vec == obj2.vec);
}

[stephenberry]

Thanks for this! I can tweak the code a bit to handle buffer bounds safety. Could you add some unit tests to the tests/eigen_test/eigen_test.cpp in this pull request?

The simple solution for writing safely is dump<'['>(b, ix);

[stephenberry]

@WanZhiQiu-ac, I guess I just collided with you, but I merged our changes.

[stephenberry]

@WanZhiQiu-ac, I'll add some more unit tests. I limited the specialization for matrices with both rows and columns being dynamic, because Glaze already supported dynamic vectors in a way that didn't need to serialize the extents.

[stephenberry]

@WanZhiQiu-ac, I've added more tests, but some are breaking, so I need to work through them.

[stephenberry]

@WanZhiQiu-ac
So, everything seems to be working now. I will note that this approach does not encode the layout (row or column major). However, transposing the data is quite expensive and it is better to serialize/deserialize in the same format. But, if you think we should encode the layout (and maybe just error if the layout doesn't match) then we can update the code. The BEVE code handles the layout, just not the JSON.

I'm going to merge this in, but make another pull request if you want to handle layout checking or transposing. But, this should now serialize/deserialize dynamic matrices.