Implement logic for feature detection piggy backing on LLVM's feature detection

[gbaraldi]

This doesn't switch or implement the new logic fully but I'm already opening for some comments if people have opinions.
The design is a bit more C++isy (this means removing c++ from the runtime is further away) and potentially less space efficient than the previous one (Not that it's very relevant).

The design switches to a clone targets logic that serializes a string like
cpu_name1;+feature1, +feature2; 0: cpu_name2; +feature3, +feature4; 0
This gets rid of the nice bitset logic that we have but hopefully LLVM can replace most of it

• CPU name based parsing
• Target choosing logic (This matches by number of features currently, do we want something nicer?)
• Extra feature parsing (dependency detection)
• Cloning logic
• Adapt LLVM pass
• Extra targets (AMDGPU?)
• C api for LLVM.jl and relateds to use (just wrap jl_get_llvm_features)
• Delete old code