CI - fix LLM benchmark workflows

[bradleyshep]

Description of Changes

Updates llm-benchmark-periodic.yml and llm-benchmark-validate-goldens.yml to new CI infrastructure.

• Switch to the current runner (spacetimedb-new-runner-2)
• Drop the dead localhost:5000 container + --privileged
• Build the local SpacetimeDB server the benchmark harness needs, and use that
  same local CLI for publishing

API and ABI breaking changes

None. CI-only.

Expected complexity level and risk

1 — workflow-only, no production code.

Testing

• Run both workflows via workflow_dispatch and confirm they pass

[cloutiertyler]

This looks good to me, but I want to ping @bfops here in regards to whether we should rebuild SpacetimeDB or just install it