Kernel Fusion TODOs

[mjguynn]

• AllocTemporary elision. If a small temporary allocation is exclusively used within a single fused kernel, that allocation should be skipped, and it should be replaced with a global variable inside the compute shader. Implementing this might be kind of ugly because the binding-to-global remapping happens at fused kernel construction time.
  • should avoid leaving gaps in the binding layout
• There's some really awful type-punning going on while translating naga::Statement::Emit during kernel fusion. Worse, it's not even guaranteed to be correct -- the process of expression remapping may turn a contiguous range into multiple disjoint ranges if we're not careful. It might be best to just halt kernel fusion if we encounter Emit.
• Fused kernels aren't added to the FFI interface due to mutability restrictions. This caused the signatures of codegen functions to change; they now take references to GPU shader modules rather than external function IDs. I'm not sure whether this is a good or bad change...
• Shaders must now have unique names.
• GpuKernel should store the actual Naga shader datastructure (in a newtyped wrapper, of course). Storing it as a string causes a bunch of redundant serialization/deserialization. The main challenge with this approach is serializing/deserializing from RON, where we want the shader's string contents. I think this can be addressed with custom serialization logic in the newtyped wrapper.
• It would be nice to have a better way of showing what kernel fusion actually does. Right now, you have to look at the generated Rust code to understand what it does.
• FuseState::finish is just really ugly. It should probably be rewritten from scratch.