fix: make sample-catalog fetches resilient to transient failures

[Yimin-Jin]

Summary

The catalog generator fetched each sample's azure.yaml / README.md from raw.githubusercontent.com and swallowed every error as return null, treating it as file does not exist. Under the ~90-request serial scan, an occasional rate-limit (429) / 5xx / network blip therefore silently dropped a real template — the same upstream commit produced 88 vs 89 templates across two runs (the C# hello-world sample vanished from one).

Changes

Replaces the silent catch-all with a resilient fetch layer:

• HttpError carries the status code so callers can distinguish outcomes.
• fetchWithRetry retries only transient failures (429 / 5xx / network); a 404 is permanent and not retried.
• Backoff uses full jitter exponential backoff and honors a server Retry-After header when present.
• Per-wait cap (FETCH_MAX_DELAY_MS) and a total backoff budget (FETCH_TOTAL_BACKOFF_BUDGET_MS) prevent a rate-limit storm from stalling past the job timeout — once exceeded the run fails fast.
• fetchAzureYaml now returns null only on a genuine 404; a transient failure that survives retries is re-thrown, so the run fails loudly instead of emitting an incomplete catalog.

Notes

• All knobs are overridable via env vars for local debugging; defaults are conservative.
• Zero new dependencies (the script is intentionally dependency-free).
• sample-catalog.json is regenerated by the sync workflow and is intentionally NOT included.

Validation

• node --check passes. Recommend a manual workflow_dispatch run before merge to confirm the full scan still succeeds.