amdgpu-waves-per-eu occupancy hints in rigid/constraint kernels are MI300X-specific - revisit for future architectures

[kevinjosephamd]

Summary

In #39, amdgpu-waves-per-eu fn_attrs were added to three hot kernels in the rigid body and constraint
solver:

Kernel	File	waves-per-eu (min,max)
kernel_step_1	genesis/engine/solvers/rigid/rigid_solver.py	3,4
kernel_step_2	genesis/engine/solvers/rigid/rigid_solver.py	1,4
func_solve_init	genesis/engine/solvers/rigid/constraint/solver.py	2,4

These values were determined by a per-kernel occupancy sweep on MI300X (CDNA3) and are not guaranteed to be optimal or neutral on other AMD GPU architectures.

What needs to be done

When new AMD GPU architectures become available, or if the workload characteristics change significantly (e.g.
different scene complexity or number of environments), the occupancy sweep should be re-run for
these three kernels to determine whether the current hardcoded values are still appropriate or need to be updated.

A longer-term alternative would be to introduce per-architecture hint selection in Quadrants (see ROCm/quadrants#11), but that is a significantly more involved change and out of scope for now.

Note: This issue serves as a reminder that these values are architecture-specific and should be revisited as
hardware evolves.