Make the NCG able to compile the RTS

[simonmar]

It is almost possible to compile the RTS with the native code generator.  What is missing is support for loops: code generated by GHC doesn't normally contain loops (or rather, loops are explicit as jumps to top-level labels), but the RTS code does contain some loops.  The register allocator in particular will need some work.  

There's nothing fundamentally difficult here, and it doesn't need to do a fantastic job, just a half-decent one.  The code to be compiled in the RTS isn't performance-critical.

[simonmar]

Now done.