Make compile_fxc accept full_stage: &CStr

[workingjubilee]

As discussed in #5812 (comment) the reason that it's acceptable to pass this String to an LPCSTR argument is because it's secretly a CStr. Make it be what it thinks it is, as otherwise, despite this having only one caller, it's not typesafe.

[workingjubilee]

• ability to write type-safe code
• ability to remember import paths

[Elabajaba]

Should I cherrypick this one for backporting as well?

[workingjubilee]

No idea how a patch for DX12 on Windows is affecting Vulkan on Linux...?

@Elabajaba Unlike the other patch, where any version of wgpu with compile_fxc and without that patch is actively inducing reads of garbage data on Windows (or segfaults, which are much less worse), this one shouldn't affect the actually-compiled programs. Rather, it affects whether compile_fxc is an unsafe fn and whether its callers must have secret magic knowledge. As this function is not public outside this codebase, the change here will have no effect on resulting programs, but will make it less likely a future change incorrectly violates invariants.

Rather, it affects whether compile_fxc is an unsafe fn and whether its callers must have secret magic knowledge.

Currently: compile_fxc is not an unsafe fn, but its callers require secret knowledge of what happens to their strings, so it should be.

With this patch: compile_fxc does not need to be an unsafe fn as its callers no longer require secret knowledge of what happens to their strings.

[workingjubilee]

"so is that a yes or no"
"idk lol I don't know what your policy is for backporting type-safety fixes?"

[teoxoy]

No idea how a patch for DX12 on Windows is affecting Vulkan on Linux...?

One of the tests is intermittently failing.