Sorry for the late reply, I've been very busy those weeks.
The current design of nphysics has not been made with GPU in mind so I guess some architectural work is to be done to make it run on GPU. Also, we would first need to improve the current (or create new) rust tools for writing/running OpenCL kernels. It has been some time since I checked the ocl crate but when I did a couple of years ago, it was still buggy and non-ergonomic. Perhaps it is ready now? Also, being able to write OpenCL kernels in Rust would be extremely useful for this kind of undertaking. There has been some attempts of DSL or to compile rust code on the GPU, but nothing more recent AFAIK.
Regarding CPU performance, the first step would be to benchmark typical simulations on all three engines, and improve each part of nphysics accordingly. I am currently working on improving performances of nalgebra (which physics uses). nphysics relies on similar algorithms as Bullet, so we should be able to at least be as performant without too much effort.
Adding support for 1D should be quite straightforward. We just have to somehow disable rotations for 1D scenarios, then most of the existing code should work fine. Of course, we may then want to do some optimizations, e.g., regarding collision detection on 1D shapes which is straightforward.
Adding support for 4d should be much harder. We would first have implement 4D rotations on nalgebra. Then, after formulating kinetic and potential energy in 4-space, we have to see how Euler-Lagrange equations of motion unroll for 4D dynamics. Perhaps some additional inertial terms appear, and higher-order tensors may become necessary (e.g. for the rotational inertia tensor?), I don't know yet. Some academic work exist for n-dimensional rigid body dynamics, but I never dug into it.
(Also, note that this year I focus mostly on nalgebra and ncollide so that they provide solid tools for future works on nphysics.)