Add new application: sbd

[william-dawson]

Depends on #74.

The branch intentionally contains #74 so the complete SBD pipeline can be
tested before the shared site fix merges. The SBD-only changes can be reviewed
at:
william-dawson/benchkit@fix-rikyu-cloud-site...add-sbd-benchkit

Summary

Add SBD selected-basis diagonalization as a BenchKit application for Rikyu
B200 GPUs and R-CCS Cloud DGX Spark.

Changes

• Add programs/sbd/build.sh with NVHPC 26.3 Thrust builds:
  • Rikyu: cc100, rank distribution, block distribution, index reordering,
    safe MPI allreduce, and NCCL.
  • DGX Spark: cc120, rank distribution, block distribution, index
    reordering, and safe MPI allreduce.
• Add programs/sbd/run.sh with one MPI rank per GPU and explicit local GPU
  binding.
• Use SBD's internal Davidson timer as the FOM instead of scheduler or shell
  wall time.
• Require SBD completion markers and validate the converged energy before
  emitting a result.
• Add Rikyu strong-scaling rows at 4, 8, and 16 B200 GPUs.
• Add a one-GPU DGX Spark row using the memory-safe H2O 1em5 input.
• Include the upstream H2O FCIDUMP/determinant inputs with revision,
  checksum, and Apache-2.0 provenance.

Validation

DGX Spark

• NVHPC 26.3, one GB10 GPU.
• Exact energy: -76.24373504205295 Ha.
• Internal Davidson FOM: 228.928853 s.
• Complete scheduler elapsed time: 4 minutes 9 seconds.
• The larger 1em7 input was rejected for this target after reaching about
  98 GB before Davidson; 1em5 completes with adequate memory headroom.

Rikyu

• NVHPC 26.3, project rkp00012, one MPI rank per B200.
• All full scaling runs converged to -76.243776776861 Ha.

GPUs	Nodes	Davidson FOM (s)	4-GPU-relative speedup
4	1	565.510	1.00x
8	2	290.222	1.95x
16	4	152.144	3.72x

A fresh validation of this BenchKit branch:

• generated the 4/8/16-GPU scheduler matrix with partition gpu and account
  rkp00012;
• rebuilt upstream SBD commit
  9481f290c2f49d4f8e5df9b0c9c87ea0f7937c2c through build.sh;
• confirmed the effective cc100 rank-distributed/index-reordered/NCCL
  CMake configuration;
• submitted the two-node/eight-GPU row through BenchKit's own
  scripts/test_submit.sh;
• completed at -76.2437767768609 Ha with a 290.281226 s internal Davidson
  FOM; and
• produced valid normalized result JSON with source provenance and the mult
  timing section.

Additional checks

• Shell syntax checks pass for build.sh and run.sh.
• git diff --check passes.
• All included input SHA-256 checksums match their recorded upstream values.