Add parallel step execution in YAML workflows

[tombee]

Problem

YAML workflows currently execute steps sequentially. Many workflows have independent steps that could run concurrently, reducing total execution time.

Concrete example from foreman: The gathering stage could run codebase analysis and documentation reading in parallel, cutting stage time in half.

Proposal

Add parallel step execution support:

steps:
  - id: parallel_analysis
    parallel:
      - id: analyze_code
        type: llm
        prompt: "Analyze the codebase structure"
      
      - id: read_docs
        type: llm  
        prompt: "Summarize the documentation"
    
  - id: combine
    type: llm
    prompt: |
      Combine findings:
      Code: {{.steps.analyze_code.response}}
      Docs: {{.steps.read_docs.response}}

General Utility

1. Multi-source analysis - Gather information from multiple sources simultaneously
2. Batch processing - Process multiple items in parallel
3. Independent validations - Run checks concurrently

Acceptance Criteria

• parallel step type that runs child steps concurrently
• All parallel steps complete before workflow continues
• Errors in any parallel step fail the parallel block (configurable)
• Step outputs accessible via normal template syntax
• Token tracking aggregates across parallel steps
• Integration test for parallel execution

Priority

Medium - Performance improvement, not blocking for MVP.

Labels

enhancement, sdk, priority:medium