Improve encapsulation of shmem and dxgi GpuMemoryBufferHandles

This CL attempts to move shared memory and DXGI GpuMemoryBufferHandles
towards a model that are easier to move into a std::variant. The
trickiest part of this is updating use of the shared memory region:
previous refactorings here ignored this by simply dynamically
dispatching depending on the type of the std::variant. While it is
possible to continue implementing this behavior, being explicit here
probably has long-term benefits for code readability...

Summary of changes:
- Extracting a shared memory regions or DXGI handle from a
  GpuMemoryBufferHandle is now a consuming operation, i.e. the
  GpuMemoryBufferHandle becomes an EMPTY_BUFFER.
- Mutable getters and setters for these two subtypes no longer exist,
  to make it impossible to have a SHARED_MEMORY_BUFFER with an invalid
  region.
- Add overloads for constructing a GpuMemoryBufferHandle from a shared
  memory region or DXGI handle.

Bug: 40584691
Change-Id: I580a85be460035e063d1bf65e3c031104cb5410c
Reviewed-on: https://chromium-review.googlesource.com/c/chromium/src/+/6381717
Reviewed-by: Frank Liberato <liberato@chromium.org>
Reviewed-by: Piotr Bialecki <bialpio@chromium.org>
Reviewed-by: Ilya Nikolaevskiy <ilnik@chromium.org>
Reviewed-by: Colin Blundell <blundell@chromium.org>
Reviewed-by: vikas soni <vikassoni@chromium.org>
Commit-Queue: Daniel Cheng <dcheng@chromium.org>
Reviewed-by: Min Chen <minch@chromium.org>
Reviewed-by: Peter McNeeley <petermcneeley@google.com>
Cr-Commit-Position: refs/heads/main@{#1445133}