Shared-texture OSR with OnAcceleratedPaint only needs exportable native
handles. Preferring kPreferMappableSharedImage forces Linux into
SCANOUT_CPU_READ_WRITE / GBM_BO_USE_LINEAR, which breaks
GPU rendering on NVIDIA. Switch to kPreferSharedImageWithNativeHandle
so Chromium can allocate SCANOUT buffers and still export DMA-BUF /
DXGI / IOSurface handles.