fix: drop inferred ingest pipeline for CSV and auto-detect delimited format

[walterra]

Summary

• Drop the ingest pipeline returned by _text_structure/find_structure when the detected format is delimited (CSV). Client-side parsing already handles CSV splitting, so the server-side pipeline is redundant and can cause conflicts.
• Expose detectedFormat from the inference result so main.js can auto-switch sourceFormat to csv when the structure is detected as delimited.
• Only set params.format when sourceFormat is not ndjson, and fall through to the actual sourceFormat value instead of hardcoding ndjson as the default.

Test plan

• Updated infer_mappings.test.js to assert no ingest pipeline is created for CSV
• Run yarn test locally (requires Docker)
• CI passes on all ES version matrix entries