Fix Rikyu and DGX Spark scheduler configuration

[william-dawson]

Problem

BenchKit's shared scheduler configuration no longer matched the live Rikyu
and R-CCS Cloud DGX Spark queues:

• Rikyu used the retired 1n1gpu partition and did not pass a project account.
• DGX Spark used the retired ng-dgx-3h partition.

Changes

• Add an optional account field to config/system.csv and expose it to
  scheduler templates.
• Configure Rikyu to use partition gpu and account rkp00012.
• Configure DGX Spark to use partition ng-dgx-m1.
• Pass the Rikyu account through matrix generation and test_submit.sh.
• Request Rikyu GPUs as the scheduler-required total --gpus=N, rather than
  the unsupported per-node form.
• Document the optional site-level scheduler account.

Systems that do not require an explicit account leave the field empty and
retain their current behavior.

Validation

• bash -n scripts/job_functions.sh scripts/matrix_generate.sh scripts/test_submit.sh
• python3 result_server/tests/check_site_config.py
• git diff --check
• Generated the matrix under Linux Bash on Rikyu and confirmed:
  • Rikyu: -p gpu --account=rkp00012
  • DGX Spark: -p ng-dgx-m1
• Used scripts/test_submit.sh sbd 2 from the dependent PR checkout to submit
  a real two-node/eight-GPU Rikyu job. Slurm accepted
  --gpus=8, the job completed successfully, and SBD converged to
  -76.2437767768609 Ha.

The SBD application PR depends on this shared site fix.