RenderFrameHostManager::current_frame_host() is expected not to be null.

This is a bit hard to guarantee, a lot of code can run inside its destructor:
#1 RenderFrameHostManager::~RenderFrameHostManager()
#2 RenderFrameHostImpl::~RenderFrameHostImpl()
#3 WebContentsImpl::RenderFrameDeleted()
#4 WebContentsObserver::RenderFrameDeleted()  (38 override excluding tests)

This is a bit unsafe. A lot of helper function assumes they can always dereference current_frame_host().

We can have rfh->frame_tree_node()->current_frame_host() == nullptr, even when rfh is the current render frame host.

============

Note: I tried several things:

1) in std::unique_ptr, exchanging ~RenderFrameHostImpl and releasing the ptr, so that current_frame_host() is still available while destructing it...
 .------------------------------------------------------
 | RenderFrameHostManager::~RenderFrameHostManager() {
 |   [...]
 |   delete render_frame_.get()
 |   render_frame_.release()
 | }
 `------------------------------------------------------
...but I think there is risk of the destructor being called inside itself.

2) Calling the observer earlier, before releasing current_frame_host().
https://chromium-review.googlesource.com/c/chromium/src/+/1206812
but it breaks a lot of tests.