Re-works how LDS regions are defined and how their total size is computed

[pm4rtx]

This PR tries to improve how LDS regions are defined in the code shared between HLSL and C/C++. Previously the size of LDS region was computed twice: in .hlsl shader file when declaring single groupshared variable and in .h file within the body of the corresponding function. The latter was necessary to represent LDS by a local array in C/C++ side. Things would have been easier if HLSL had supported declaring groupshared variable within a function, but it doesn't support that. So the first approach existed for some time to somehow address what HLSL can't do, but this approach was prone to errors when sizes of groupshared memory became out of sync. To address this, this PR starts defining partition (groups) of LDS variables by respective macro lists (usually, right in front of functions that use them) that are expanded twice: outside of the function to compute the total size of required LDS region, and the second time within function body to compute offsets relative to the start of LDS (on HLSL side), and to define local variables replacing global LDS memory (on C/C++ side). This approach requires to define all macro/constants that affect LDS size before including zstdgpu_shaders.h where LDS partition macro are defined.

[coopp]

Nice!. This looks good to me.