Lukasz Stafiniak

The semantics is problematic, likely broken. Also for to-power-of `**` with `Compose`. Maybe should be prohibited. Expanding as matrix pseudo-inverse would be overkill...

Factor out helper functions.

Also, issue warnings when broadcasting? Study the following and apply the lessons learned. * [Haliax: Named Tensors in JAX](https://colab.research.google.com/drive/1TiTcQQ4V5mopbgCu1SVl-oqJtXn7rFnC) * [Tensor Considered Harmful: NamedTensor](https://nlp.seas.harvard.edu/NamedTensor)

Low priority cosmetic problem.

Inconvenience / confusing: when printing assignment update, the RHS will contain the updated value (since we just print the current value of the cell).

Synchronizing the context once all H to D, resp. D to H, copying of a session update is issued.

``` * Allocating * excessive amounts of memory with ::cuMemAllocHost() may degrade system * performance, since it reduces the amount of memory available to the system * for paging. As...

For things like sizes in bytes, `grid_dim_x` etc. Currently I'm suggesting `int` everywhere for convenience, except in `kernel_param.Size_t`.

AVX / AVX2 intrinsics are available in most recent `libgccjit` but not ported in the binding. Inspired by [ggml](https://github.com/ggerganov/ggml).

This will allow using dedicated operations on the backends.