Investigate inlining the fast path of jemalloc

[pcwalton]

jemalloc seems to have the ability to inline the fast path of allocation. It would be awesome to do cross-language inlining so that we can make allocation equivalent to pulling a pointer off the free list in many places.

This would help get Servo flow allocation performance on par with WebKit/Blink's.

[pcwalton]

cc @thestinger

[thestinger]

The fast path is currently far too large for inlining it to work out well. The function call itself is a small part of the cost. It needs to handle a lot of special cases like coping with thread-local data destructors. It might make sense to poke some holes in the API based on guarantees on the Rust end, but I think we should deal with the lower-hanging fruit before considering that. I think there's at least 20-30% more to gain inside jemalloc itself from thread cache improvements.