Fix very rare bug around JIT code cache collection.

Fix very rare bug around JIT code cache collection.

The bug is the following:
1) JIT thread: We start a code cache collection.
2) JIT thread: We mark all code that is in the call stack of all
   threads.
3) Mutator thread: after marking its stack, resumes and does call
   that pushes JIT compiled code to the call stack.
4) Mutator thread: deoptimizes compiled code of ArtMethod Foo,
   and therefore updates the entry point of Foo through
   JitCodeCache::InvalidateCompiledCodeFor.
   (Note that updating the entrypoint could also be done through
   instrumentation).
5) JIT thread: Call JitCodeCache::RemoveUnusedAndUnmarkedCode.
   The method used to remove entries that were not entrypoints.
   It sees the compiled code for Foo but that is not an entrypoint
   anymore, so deletes it.
6) Mutator thread problem: it now has compiled code in its call
   stack that is deleted.

If it's only one mutator thread, we only hit a DCHECK when walking
the stack, as we are now seeing an invalid pc. The deoptimization
will longjmp to the caller of that invalid entry anyway.

However, if multiple mutator threads are involved, one thread
might invalidate the compiled code while the other is still
running it. And we end up deleting code that is in the call
stack of a thread, and we will crash.

The fix is to mark entrypoints before marking call stacks,
so that anything a thread might jump to is marked and kept.

bug:27424509
bug:23128949
bug:26846185

Change-Id: I07cd08cedd96b9900629f7535e95404f622104ea