Align rasterizer state, readback paths, cull mode, triangle winding order across backends

[EmilioLaiso]

Summary

Make the three backends use the same image coordinate system, same culling mode and same SV_IsFrontFace polarity.

The framework had the following setup:

1. Front-face polarity differed.
   DX/MTL Y-flip clip→framebuffer at the viewport; Vulkan with positive-height viewport doesn't.
   Same HLSL triangle ended up with opposite framebuffer winding on VK vs DX/MTL, so SV_IsFrontFace gave opposite answers for the same input. This surfaced while writing a graphics-stage system-values test that needed SV_IsFrontFace to agree across backends.

2. Visual PNG output was aligned by two compensating flips.
   DX and MTL each had a row-reversal on readback documented as "render target origin is top-left, while our image writer expects bottom-left".
   The PNG writer reversed rows again on the way out.
   VK had neither flip, but VK's no-Y-flip viewport meant its framebuffer was already "upside down" relative to DX/MTL, so its single PNG flip happened to produce a visually identical image. Two-flips-cancel on DX/MTL, no-flip-needed on VK, by accident.
   Adding any new behavior on top of this (e.g., the negative-height-viewport change in the earlier stage of this PR that kicked off the investigation) immediately broke the cancellation and produced visually-flipped outputs.

PR

Pick HLSL/DX semantics as canonical:

• Clip-space Y up
• Framebuffer Y down
• CCW = front

Untangled the asymmetry by removing the readback flips on DX/MTL and the PNG writer's row reversal together with the VK viewport flip, so all three places change as one atomic step. After that, every backend produces the same Y-down memory layout end-to-end.

Exposed shared copyFromTexture function that all backends can call.
If #1113 goes in main, which introduces Map/Unmap as functions on the Buffer type, the readback path can be generalized further.

Golden Images

All graphics golden images still work as expected. The resulting Y-order is maintained.
The Mandelbrot.test image required updating as it's a compute path. Here's the PR: llvm/offload-golden-images#6

[manon-traverse]

I'm fine with either winding mode, but this feels like something we should add to the documentation.
In the future I can imagine we can specify this in the test, but we can then use what we pick now as a default.

[manon-traverse]

The CI is failing on a Mandelbrot test. @EmilioLaiso Could you check if it uses a golden image and if that is still correct? Perhaps there is still a bug in the texture readbacks?

[MarijnS95]

Bit odd that copyFromTexture() takes a raw CPU pointer and not a Texture object, but it makes sense from the perspective that it's a mapped readback pointer that was the target of a Texture-copy operation. Not sure I have a better name in mind though, so let's leave it if you don't have anything better either.

(copyFromReadbackTexture()?)