As no one will ever try to allocate a Thread directly, we can just make
Thread a proxy of ThreadImpl by reinterpret_casting between both types.
This allows to not mention ThreadImpl in the header at all, and to
define it in an anonymous namespace in the implementation, allowing for
more optimizations.
It also saves runtime, as no memory load is needed to forward from
Thread to ThreadImpl, and we do not need to allocate ThreadImpl objects
on the heap.

R=titzer@chromium.org

Review-Url: https://codereview.chromium.org/2643443002
Cr-Commit-Position: refs/heads/master@{#42450}