-flto -fsanitize=cfi cause issues on the CUDA CI node

[RossBrunton]

Enabling CFI sanitisation causes the linker to randomly fail when ran on the nodes used to build the CUDA build. Note this is not due to CUDA itself - building the native cpu adapter similarly fails.

Outcomes I've noticed are:

• Failing to find ur_getenv, even though it should be available in one of the .o files linked to the binary.
• Crashing in LLVM when doing LTO.
• Reporting invalid debug information (seems to be fixed by setting the debug info version to DWARF4).

This should be investigated and fixed.

[RossBrunton]

After many, many days of tracking it down, this seems to be a configuration issue on the build nodes.

Some of them have llvm-linker-tools-13 installed as well as llvm-linker-tools-14, which provides the gold plugin used for lto. Due to the way plugins are loaded, version 13 has priority over 14, but doesn't seem to be able to handle llvm 14 IR in all cases. This means ar silently fails to produce an archive containing all the symbols, which only raises an actual error when the linker finds out that it doesn't have all the symbols it needs.