Consider rebuilding our codegen framework using MLIR

[thesamovar]

See https://mlir.llvm.org/

The documentation is a bit sparse at the moment, but take a look at some of the tutorials to get an idea.

Basically, it looks like it's about being a standard way to do translation between languages in a slightly more intelligent way than our templating approach. Might have a slightly higher initial overhead but once it was done we'd get a ton of options for free and wouldn't have to support every language/target hardware ourselves.

It still looks a little early days to be a good route right now, but I think we should keep the option open for the future.