Switch to LLVM's tailcc calling convention by passing arguments without env

[marvinborner]

It's not clear how much this change influences the performance of the LLVM backend; we should write/merge the new benchmarking infrastructure first before merging this.

[marvinborner]

I was able to fix the stack overflows after some amount of debugging. However, certain tests now take >10min or do not seem to finish at all (EDIT: before the LLVM opt pass, after optimization the slow-down seems to be around 5-10x). I'm unsure whether this can be easily fixed or is implied by our use of tailcc.

IO tests still segfault, but not with stack overflows but NPE.

[b-studios]

To document what @phischu told me: we only have three locations where we really NEED tailcc. All of which are generated and not in hand-written LLVM code:

1. jumps
2. return
3. invoke

Since we are in CPS those need to be tail, for all others we do not care necessarily.