docs(evaluate): Ablations and iterations page — comparing runs systematically

[sephmard]

Part of #1436 — Evaluate section restructure.

There is no page explaining how to structure ablation studies or model iteration comparisons using NeMo Gym — a key use case for the Frontier Eval workflow.

Tasks

• Create fern/versions/latest/pages/evaluation/ablations.mdx
• Cover:
  • What an ablation is in the Gym context (vary one dimension — model, harness, prompt, verifier — while holding others fixed)
  • Recommended run structure for a clean comparison (fixed seed, repeat count, sampling settings)
  • Naming and organizing runs for later analysis
  • How ablation outputs feed into BLADE analysis (link to Benchmark Analysis / BLADE)
  • Iteration pattern: baseline → intervention → re-run → compare aggregate metrics
• Add navigation card in evaluation index
• fern check passes