gfxstream: Fix crash and synchronization in GPU scaled readback

[GOOG-sergiu]

When Android Studio or a client requests a scaled display frame (via gRPC streamScreenshot), readColorBufferPixelsScaled delegates to readColorBufferPixelsScaledGpu. This previously caused an immediate crash (SIGSEGV / exit code 139) on the host due to several Vulkan synchronization and memory violations:

• Missing layout transition and memory barrier: sourceCbInfo->image was sampled in CompositorVk::drawImage() as VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL without transitioning from its current layout (e.g. COLOR_ATTACHMENT_OPTIMAL or GENERAL). On host drivers, sampling an attachment layout without a barrier triggers GPU MMU faults. Added a pre-draw pipeline barrier transitioning to SHADER_READ_ONLY_OPTIMAL and a post-draw barrier restoring the original layout.
• Missing outbound subpass dependency: tempRenderPass only specified an inbound dependency (EXTERNAL -> 0), causing vkCmdCopyImageToBuffer to execute concurrently with active color attachment writes. Added an outbound dependency (0 -> EXTERNAL) to flush color attachment writes before the transfer read.
• Missing staging buffer size check: validated requiredStagingSize against mStaging.mAllocationSize.
• Hard aborts on transient errors: replaced VK_CHECK with graceful error logging, resource cleanup, and queue idling on fence failure.
• Added graceful fallback to readColorBufferPixelsScaledCpu if readColorBufferPixelsScaledGpu fails.

Bug: 553490628
Test: bazel test //host/vulkan:vk_common_operations_tests //host/vulkan:gfxstream_compositorvk_tests
Change-Id: Icb7889bb7bbdc8dbe3d3ffdae0ba2899dfd85a2f