Deprecate #[spirv(block)] and auto-wrap in "interface blocks" instead.

[eddyb]

As per Vulkan and OpenGL requirements:

• PushConstant, Uniform and StorageBufer interface variables must have a Block-decorated OpTypeStruct type
  • this "struct" type is called an "interface block" and serves mostly as a grouping construct
  • as per Make #[spirv(block)] decoration force ABI type into Struct #391 (comment), having a struct (with some fields) inside an "interface block" shouldn't differ from having those same fields in the "interface block" itself
  • after this PR, we automatically generate a single-field "interface block" for these storage classes, instead of requiring the user to declare a struct and annotate it with #[spirv(block)]
• Input/Output can also use "interface blocks", but it's not mandatory
  • we'd probably want to think a bit more about this once/if we have usecases in mind

Most of the commits in this PR are small refactors, leading up to the last commit, which is the actual change.

Because of #[spirv(block)] being deprecated instead of outright removed, this isn't a breaking change, except for slices, where you have to remove the struct wrapping the [T] that you previously had to add - it's now just:

#[spirv(storage_buffer, descriptor_set = 0, binding = 0)] slice: &mut [f32],

The auto-generated "interface block" also allows simpler uniforms like this to "just work":

#[spirv(uniform, descriptor_set = 1, binding = 0)] u_view_proj: &Mat4,

Closes #391.

EDIT: temporarily rebased on top of #577 to unbreak CI.

[fu5ha]

Very nice :D

[eddyb]

@khyperia @XAMPPRocky For the record I think this could be avoided in the future if GitHub has an option to require reviews from everyone who was added as a reviewer.

EDIT: can't find such an option in the branch protection settings of my own repos, welp.

[XAMPPRocky]

What happens if there's multiple interface blocks in the function? Won't this point them to the same offset?

[eddyb]

This is just what you would get from this kind of struct:

#[spirv(block)]
struct InterfaceBlock<T> {
    inner: T,
}

So this isn't a regression, and I believe offsets are per interface variable.

There's an unrelated issue that we allow multiple push constants and multiple other bindings with the same (descriptor_set, binding), that's the only situations I can think of in which anything would overlap.

[XAMPPRocky]

.debug_struct("InterfaceBlock")

[khyperia]

LGTM

[khyperia]

12 levels of indentation makes me sad :( would be nice to split this out

[eddyb]

I guess the newly introduced use of Builder means that we can interleave things a bit more and not have to separate them out so much - e.g. declare_interface_global_for_param could also return an OperandValue for the call args.