To improve performance of parking, keep the thread state in an atomic
variable instead of protecting it with a mutex.

However the mutex was used e.g. to force Unpark() to block while the
safepoint operation was still running. Therefore the safepoint algorithm
has to change as well.

Park() and Unpark() use CAS operation to transition the state.
Safepoint() uses a relaxed load for checking whether a safepoint was
requested. Since Safepoint(), Park() and Unpark() all have a slow path,
there is no need for busy-waiting on the main thread.

We need two more ThreadStates:
* SafepointRequested: This state is set by GlobalSafepoint to force
    Running threads into the slow path on Safepoint() and Park(). This
    state also replaces the separate atomic<bool> safepoint_requested_
    field we used before.
* ParkedSafepoint: This state is set by GlobalSafepoint as well to force
    parked threads into the slow path on Unpark().

When stopping all threads, GlobalSafepoint transitions states from
Running --> SafepointRequested and Parked --> ParkedSafepoint to force
the slow path for all three methods. After performing the transition
for each thread we know the exact number of Running threads and wait
until each of them either reached a safepoint or parked itself.

Design doc: https://docs.google.com/document/d/1p9klWyqT_AScAnK_PdHZTcNhZGzoBiYWPkUciIh2C58/edit?usp=sharing

Bug: chromium:1177144, v8:10315
Change-Id: I8697da915c7d18e2fb941f1bedf6181226408feb
Reviewed-on: https://chromium-review.googlesource.com/c/v8/v8/+/2704075Reviewed-by: Ulan Degenbaev <ulan@chromium.org>
Reviewed-by: Maya Lekova <mslekova@chromium.org>
Commit-Queue: Dominik Inführ <dinfuehr@chromium.org>
Cr-Commit-Position: refs/heads/master@{#73089}