JSON with a UTF-8 BOM is misdetected as CSV by sqlite-utils memory #940

On main (9c66baf, sqlite-utils 4.2.1), sqlite-utils memory misdetects a UTF-8 JSON file as CSV when it starts with a byte order mark (BOM). The JSON reader already accepts the same bytes when the format is explicit.

With Python 3.12 on macOS, run these commands in a temporary directory:

python -c 'from pathlib import Path; Path("items.json").write_bytes(b"\xef\xbb\xbf" + b"[{\"id\": 1, \"name\": \"Cleo\"}]")'
sqlite-utils memory items.json 'select id, name from items'
sqlite-utils memory items.json:json 'select id, name from items'

Actual: autodetection exits with Error: no such table: items. The explicitly typed command returns [{"id": 1, "name": "Cleo"}].

Expected: both commands return the same rows. rows_from_file() checks the first bytes for [ or { before accounting for a BOM. The fix can normalize only the detection sample and leave the actual input untouched for json.load().

I have a small patch with offline regression coverage. Codex assisted with the source review, reproduction, and patch.