Propose an initial graph refactoring framework for shader generation

[jstone-lucasfilm]

This changelist proposes an initial graph refactoring framework for MaterialX shader generation, replacing hardcoded graph optimizations with a composable set of named refactoring passes. The following specific changes are included:

• Add a ShaderGraphRefactor base class, with three concrete subclasses for node elision, premultiplied BSDF add (valuable in hardware languages), and layer-over-mix distribution (valuable in MDL).
• Add a virtual setDefaultOptions method to ShaderGenerator, enabling each subclass to set target-appropriate defaults at GenContext construction time.
• Add a MIX classification bit to ShaderNode for efficient mix node identification.
• Remove hand-authored GLSL and MDL graph overrides for open_pbr_surface and standard_surface, as these refactoring steps are now generated automatically.

[jstone-lucasfilm]

This proposal will require significant discussion from the MaterialX community, and I'm initially CC'ing @niklasharrysson, @ld-kerley, @ppenenko, and @jreichel-nvidia for their expertise and review.

This changelist directly addresses issue #2480, but it goes significantly further than this, providing a composable system of graph refactoring passes that handles constant/dot node elision, BSDF mix pre-multiplication for hardware shading languages, layer-over-mix distribution for MDL, and potentially far more in the future.

Let me know your thoughts, and I'd prefer that we take our time with this proposal, rather than rushing it into a release of MaterialX.

[jstone-lucasfilm]

Here I'll share renders and performance numbers for common OpenPBR Surface materials, and the general pattern is that the current refactoring PR actually provides better real-time shading performance than the original manually-optimized graphs, which was not the case for earlier automated optimization proposals:

OpenPBR Surface Carpaint, Manual Optimization: 8 ms/frame

OpenPBR Surface Carpaint, Current PR: 7ms/frame

OpenPBR Surface Glass, Manual Optimization: 12ms/frame

OpenPBR Surface Glass, Current PR: 11ms/frame

OpenPBR Surface Pearl, Manual Optimization: 13ms/frame

OpenPBR Surface Pearl, Current PR: 12ms/frame

[ppenenko]

Looks great @jstone-lucasfilm! I think it's important to clarify that these changes would not conflict with my current WIP on #2680 - even though I initially experimented at the shader graph optimization/rewriting stage as well, my code ended up moving to the shader graph construction stage. So, the two optimizations should be complimentary in the end.

[ppenenko]

I would also suggest the infrastructure we discussed recently as an automated way to measure the performance impact of these changes with more precision and test against correctness regressions, namely:

• Extended tracing instrumentation for shader codegen and rendering #2820 and GPU async timing instrumentation via MX_TRACE_ASYNC #2824: nanosecond-precision instrumentation instead of truncated milliseconds in MaterialXView; GPU timer queries with proper CPU/GPU synchronization; checking the impact on codegen and compilation times;
• Add envSampleCount and framesPerMaterial test suite options #2821: framesPerMaterial to get more frame duration samples from MaterialXTest for averaging; envSampleCount for conditions more representative of interactive rendering and not as dominated by texture sampling;
• Python comparison scripts for test suite output (diff_test_runs) #2822: automated perceptual image comparison.