Port sparse inequality constraint Jacobian to GPU

[pelesh]

Merge request type

• New feature
• Resolves bug
• Documentation
• Other

Relates to

• OPFLOW
• SOPFLOW
• SCOPFLOW
• TCOPFLOW
• CMake build system
• Spack configuration
• Manual
• Web docs
• Other

This MR updates

• Header files
• Source code
• CMake build system
• Spack configuration
• Web docs
• Manual
• Other

Summary

Replace PETSc-based inequality Jacobian with GPU RAJA kernels

Move the inequality constraint Jacobian computation for the HiOp sparse
GPU solver entirely to the device, eliminating the per-iteration host
back and forth (copy to host, PETSc compute, MatGetRow extraction, values
copy back to device). Elimiate PETSc use from this part of the code.

Three RAJA kernels now compute directly into device memory:

• Generator set-point constraints (AGC)
• Voltage-reactive-power bounds (FIXED_WITHIN_QBOUNDS)
• Line flow limits (Sf^2/St^2 derivatives + slack variables)

Supporting changes:

• Analytical NNZ counting replaces PETSc MatGetInfo at solver setup
• New device-side parameter fields (apf, vs, xpdevidx, xslackidx,
  bus-to-gen mapping) added to *ParamsRajaHiop structs
• Sparse position indices assigned at model setup for all three
  contribution types

Includes validation test (test_ineqjac_gpu) that solves with IPOPT,
then compares PETSc and GPU Jacobian values at the converged solution.
Optional -benchmark flag for performance comparison.

Made-with: Cursor

[nkoukpaizan]

Just a couple more minor suggestions. Looks good otherwise!

Benchmark results are promising. It would be good to see how the GPU kernel compares to a purely CPU evaluation (i.e., what portion of the PETSc path was compute versus data movement).

case_ACTIVSg200.m

=== Performance Benchmark (1000 iterations) ===
  PETSc path (compute + MatGetRow + copy): 85.9155 us/iter
  GPU path (RAJA kernels, no copies):      14.2352 us/iter
=== End Benchmark ===

case_ACTIVSg2000.m

=== Performance Benchmark (1000 iterations) ===
  PETSc path (compute + MatGetRow + copy): 670.548 us/iter
  GPU path (RAJA kernels, no copies):      15.0518 us/iter
=== End Benchmark ===

case_ACTIVSg10k.m

=== Performance Benchmark (1000 iterations) ===
  PETSc path (compute + MatGetRow + copy): 1975.1 us/iter
  GPU path (RAJA kernels, no copies):      15.803 us/iter
=== End Benchmark ===

[nkoukpaizan]

Looks good to me! Tests are passing on the HIP and CUDA backends.