Implement specialization constant support in numthreads / local_size

[juliusikkala]

Adds support for using SPIR-V specialization constants in numthreads:

[vk::constant_id(0)] const int WG_SIZE = 8;

[shader("compute")]
[numthreads(1, WG_SIZE, 1)]
void main(uint3 tid : SV_DispatchThreadID)
{
}

Also adds support in GLSL mode:

#version 460

layout(constant_id = 0) const uint size = 8;
// The standard local_size_x_id is supported, but also local_size_x = SPECCONSTNAME.
layout(local_size_x_id = 0, local_size_y = size) in;

void main()
{
}

Similar to the local_size_x_id feature from GLSL, this allows users to generate shader permutations with different work group shapes / thread counts after the shader has already been compiled to SPIR-V.

Currently, this works with the SPIR-V (using LocalSizeId execution mode) and GLSL (using local_size_{x,y,z}_id) targets. I'm not sure if there even is anything that could be done for the other targets, other than to just present diagnostics about this feature not being supported there.

[juliusikkala]

I've now tested this in a real Vulkan program in a couple of different contexts, and this feature seems to now work correctly.

[csyonghe]

Can you add a test for this feature?

[csyonghe]

Looks good to me. I wish there is a cleaner way to add this feature, but what is implemented in this PR is probably the easiest path forward. If I understand the code correctly, this implementation won't allow something like [numthreads(specConst+1, 1, 2)], right?

[juliusikkala]

It currently doesn't allow doing math on the specialization constant, because I'm fairly sure that that can't be expressed in SPIR-V. I'll fix that compiler warning later today, I've been developing on Linux only so I guess MSVC errors are to be expected.