WIP: add an L3 -> LLVM IR translation pass

[zerbina]

Add a pass that translates L3 code into LLVM IR, making it possible to compile source language programs and NimSkull program's compiled via skully to native executables via LLVM.

At the moment, the pass is biased towards code produced by skully.

Instead of using the LLVM API to create the LLVM IR and compile it into an executable, the pass simply emits textual LLVM IR, which has a number of benefits:

• no bindings to the LLVM API are needed
• phy doesn't depend on a dev version of LLVM; LLVM doesn't have to be built from source to compile phy
• the textual IR is much more stable than the LLVM API

In the future, the pass could move to emit LLVM bitcode, which would speed up the pass and produce smaller artifacts.

---

To-Do

• fix the upstream (i.e., NimSkull) bugs that make skully produce incorrect code
• move the skully fixes into a separate PR
• move the Ptr type introduction to a separate PR
• split Conv into multiple operators (separate PR)
• properly implement exception handling
• implement Select support
• split up runtime.c
• test skully + LLVM pass in the CI
• use the L25 IL as input
• write a proper commit message

[zerbina]

Locally, I'm able to compile phy.nim into a working native executable, like so:

> skully phy/phy.nim phy.txt
> phy --source:L25 --target:llvm c phy.txt
> llvm-as -o phy.bc out.ll
> llc -filetype=native -o phy.o phy.bc
> clang phy.o runtime.c 

It's also possible to combine the LLVM steps by using (stock) clang, but this routinely resulted in clang crashing, at least on my system.

The above is not reproducible with a stock NimSkull yet, as there are multiple bugs in the NimSkull compiler where incorrectly typed MIR code is emitted. Since skully doesn't perform any type correction (nor should it), the incorrectly typed MIR leads to incorrectly typed L3 code, and since the LLVM IR is quite strongly typed (when compared to, e.g., the Phy VM bytecode), this presently results in llvm-as rejecting the input.

---

This PR will most likely stay open for quite some time, as there's no rush to finish it. It's also not clear whether it should be merged at all, or whether it's better for the LLVM pass to be implemented in a separate repository.

I mainly wanted to experiment with how easy/possible it is to implement a to-LLVM translation pass with the current compiler.

[zerbina]

I've merged the PR with upstream and also added a test for the code generator to the CI.

There's still one issue remaining in the NimSkull compiler, namely, that the pred and succ translation in mirgen produces incorrect MIR code, which shows up here as a type error from the LLVM assembler.

Separately, I've concluded that the IL Select was a misfeature. I'll replace it with something akin to a branch-with-table, which is much simpler to implement, not only for the LLVM code generator, but also the VM code generator.