Add an AWS Batch workflow alongside NERSC

[joglekara]

Summary

• add an aws-batch-run skill alongside nersc-workflow
• use the deployed continuum-infra/sim-runner bundle rather than the retired adept.cloud client
• document queue and job-definition selection, dry-run-first submission, CloudWatch monitoring, exact-job termination, and MLflow verification
• install the skill for both Claude Code and Codex from bootstrap-local.sh

This is a fresh, current replacement for closed #8. The runtime and infrastructure already live in continuum-infra; this PR adds only the agent-facing workflow.

Notable updates from #8:

• gpu now has an on-demand fallback plus an explicit gpu-ondemand queue
• insufficient-capacity jobs are cancelled after 30 minutes with a concrete status reason
• the shared Batch client was intentionally removed from adept, so the skill resolves the private infra checkout directly
• production runs require a clean pushed commit in addition to recording the content-addressed bundle SHA

Validation

• skill-creator quick_validate.py: passed
• bash -n scripts/bootstrap-local.sh: passed
• git diff --check: passed
• exact documented dry run against kinetic-srs: passed
  • 705 files
  • 0.7 MB compressed bundle
  • deterministic SHA-256 reported
  • no S3 upload and no Batch submission