Root arguments that outlive their signature, and depth clears that round

What changed

DXGI (the Windows layer that lists GPUs and displays) now reports real displays and display modes, and it reports how much GPU memory the process uses. After that I worked through small conformance failures, one at a time.

• Draws now bind every vertex-buffer slot the input layout reads. Empty slots read zeros instead of stale addresses.
• A pipeline that declares no depth format still applies its depth and stencil tests to a bound depth buffer.
• Changing the root signature keeps the root arguments already set. A root signature is D3D12's description of which constants and resource tables a shader gets. ClearState is implemented and zeroes them.
• Shader resource arrays are checked against the ranges of the root signature.
• Descriptor tables that point outside the bound heap read the heap they point into.

The interesting part

Two fixes copy what native drivers do, even though the D3D12 documentation says little or nothing about it.

Root arguments. vkd3d-proton's test_uninit_root_parameters sets constants, switches to a second, compatible root signature, and dispatches. It expects the old constants to still be there. The test's comment says native drivers do not clear anything on a root-signature change. I had been zeroing the arguments on every change. Now only Reset and ClearState zero them.

Depth clears. Clearing a 16-bit depth buffer to 0.5 gave 32767 on Metal. D3D12 converts floats to 16-bit normalized values by rounding, which gives 32768, and test_depth_stencil_sampling compares against exactly that. Metal appears to truncate. The layer now nudges the clear value so that truncation lands on the rounded result. A new API test checks five values, including 0.37 and 0.9999, both for a plain clear and for a clear with a rect, which the layer draws. The review asked for the second case.

In one case the two test suites disagree. After IASetVertexBuffers(0, n, NULL), vkd3d expects the slots to be unbound. vkd3d-proton expects the call to be ignored. vkd3d's test skips NVIDIA as "broken", which suggests this is a vendor difference and not a rule. I kept vkd3d's behaviour and recorded the conflict.

Numbers

• Full run: 495 pass, 93 fail, 190 skip of 778, 0 newly failing.
• Previous full run: 482 / 102 / 190 of 774. Four of the 13 newly passing tests are new API tests.
• Failed checks: 3,861 -> 3,843.