[techomancer]

A common wgpu layer (src/wgpu) for host-GPU rasterisation, and the VPro (Odyssey) raster unit built on it, word for word with the CPU pipeline.

• The board draws on the GPU raster unit with board memory mirrored on the CPU; dirty maps both ways, region readbacks, no waits on texture and image uploads (blit records, mip levels and scratch images as regions).
• The texture unit runs on the GPU.
• Batches submit as the board idles or after 4 ms; display swaps wait on stream positions, so a flip shows only drawing that has completed.
• Window-id bytes written by GPU blits are also stored in the mirror, since the display front end reads them without waiting.
• The VPro display composes on the GPU.
• New window renderer: renderer = "wgpu" (or --renderer wgpu). VPro frames stay on the device, nothing is read back unless a screenshot or VNC asks. Newport's compositor runs as a WGSL kernel. GL stays the default; iris-gui is unchanged.
• Testing: the CPU and GPU units agree on recorded sessions (new test), and on the guest GL suites, visual sweeps, piglit and seven X visual modes. New guest suite drivers in test/gltest (glsuite.sh, xsweep.sh, glsuite-compare.sh) and replay knobs (VPRO_REPLAY_WRITES, VPRO_REPLAY_TICK).

Squashed from the wgpu-raster branch.