Make import-wikidata WDQS queries resilient with retries and backoff

[arnekaiser]

WDQS (Wikidata Query Service) requests can fail transiently with HTTP 429/5xx responses, connection errors, timeouts, or truncated/garbled JSON bodies (e.g. JSONDecodeError at char 12881888). Previously a single such failure aborted the entire import run, losing all progress.

Fixes #468

Changes to bin/import-wikidata:

• wd_query() rewritten with a retry loop (up to WDQS_MAX_RETRIES=5) using exponential backoff (WDQS_BACKOFF_BASE=5s -> 5,10,20,40,80s) and honoring the Retry-After header. Handles connection errors, timeouts, retryable HTTP statuses (429,500,502,503,504), unexpected Content-Type, empty response bodies, and unparseable JSON.
• Add _backoff_sleep() helper computing the retry delay.
• Add an explicit request timeout (WDQS_TIMEOUT=120s).
• Reduce batch sizes to be gentler on WDQS: create_ids_from_wdqs 5000->1000, resolve_redirects 1000->500.
• Add a 1s inter-batch delay (WDQS_INTER_BATCH_DELAY) between consecutive WDQS batches to respect rate limits.
• Factor out save_cache() and persist the cache on permanent WDQS failure (in main) so a re-run can skip already-retrieved IDs.
• Save cache whenever use_wdqs and new_ids are set (not only when cache is non-None).