JVM bulk loads: the subnormal check walks every row when the text contains "e-3" (UUIDs, hex ids) #182

Since #180, the JVM bulk JSON paths (insert_many, graph batches) check each chunk for a subnormal double by searching the JSON text for "e-3" and, on a hit, walking every row in Python (type_conversion._dumps_checked). "e-3" is common in ordinary text: hex ids and UUIDs (...4fe-3a...), names like name-3. A load whose rows hold such strings takes the per-row Python walk on every chunk, for a value (a subnormal double) that almost never occurs.

Fix: search for the exponent json.dumps writes for a subnormal (e-308 to e-329 after a digit) with a compiled pattern, which costs the same C-speed scan and no longer matches ordinary text. A test checks that UUID-like and name-3 text never reaches the walk, and that a subnormal still takes the per-row path.