Implement custom UnmarshalYAML methods for WorkflowFile, Job, and Step
structs to avoid reflection overhead and eliminate redundant node traversal.

Key optimizations:
- Single-pass extraction of struct fields and JobOrder in WorkflowFile
- Direct field assignment instead of reflection-based Decode calls
- Pre-allocated slices and maps with exact capacity
- Simplified LoadWorkflowFile by removing parseWorkflowFromNode and
  extractJobOrderFromNode helper functions

Benchmark results (10 jobs × 5 steps each):
- Speed: 774,299 ns/op → 693,108 ns/op (10.5% faster)
- Memory: 206,857 B/op → 194,120 B/op (6.2% reduction)
- Allocations: 3,837 → 3,109 (19.0% reduction)

Code simplification: load.go reduced from 84 to 28 lines.