pinocchio VS stdBLAS

Open-loop walking + physics: (e.g. ZMP-based) one thing you can do (that does not imply adding an IMU to Kayra and going for feedback) is include some walking physics into your walking trajectory generation. To give you an idea, I've written a tutorial on doing that in Python a couple of years ago. The libraries in this tutorial are deprecated now, but I'm working on an equivalent stack at github.com/tasts-robots using more durable software like pinocchio. The libs are still WIP but if you are interested in exploring that dev path (i.e. making a ZMP-based trajectory generator for Kayra) I can support you and help adapt them (because I want that SW to be useful for a maximum number of people).

Drake and Pinocchio both use the Python interface to Meshcat as an option for visualizing robots, and they can both do the forward kinematics and have visualizer components that place the visual meshes properly relative to the computed frames, but both of those are full-fledged and complex kinematics and dynamics libraries for contact-rich interaction (and more), so the API calls to get a model set up are doing a lot more than providing a visual model you can feed joint angles to.

Pinocchio sounds like exactly what you need: https://github.com/stack-of-tasks/pinocchio

You can have a look at https://github.com/kokkos/stdBLAS (it's an implementation of the proposed linear algebra extensions for future C++).

It has the same BSD license as Kokkos: https://github.com/kokkos/stdBLAS/blob/main/LICENSE

The BLAS and LAPACK have C interfaces, and they generally don't allocate memory unless you ask them. (Some BLAS implementations allocate internal storage for rearranging matrix data.) There are a few C++ BLAS wrappers; here is one of them.

The second one is stdBLAS by kokkos. It doesn't even use cuBLAS but it's a reference implementation of P1673. This paper describe what could be a blas API in the std. I would recommend trying to do something equivalent on your side.