radeonsi: optimize: gl_FrontFacing, shader key updates, Viewperf on Gfx11, PS inputs

These changes help Viewperf. Summary:

• gl_FrontFacing is replaced with a true/false constant for points and lines (always true) and when face culling is enabled (front faces culled -> false, back faces culled -> true)
• Shader key updates are rewritten/optimized to call si_update_shaders less.
• small streamout and vertex element changes
• Gfx11: Performance tuning for Viewperf (use Wave64, disable shader profiles)
• Fully eliminate PS inputs from register settings if uniform inlining or colormask eliminated them in NIR.
• DSA and rasterizer states are converted to tracked registers using custom emit functions.
• The guardband state is updated much less often.