Reintroduce jemalloc

[brson]

It would still be great to have our own, awesome allocator. I'd suggest building jemalloc in the mode where it doesn't attempt to replace the malloc and free functions but instead provides them under their own name, and modify std to use jemalloc directly.

[huonw]

Won't using different symbol names interfere with LLVM's libcall optimisations? (Of course, the net benefit of using jemalloc might still be positive.)

[brson]

@huonw Maybe, but if so we can fix llvm.