Keep multipart CSV rows intact when a part lacks a trailing newline

[DMZ22]

Closes #1778.

The multipart byte stream concatenates the parts line by line, so a CSV part whose last line has no trailing newline glued that row onto the first data row of the next part — chunk1.csv = id,name\n1,english plus chunk2.csv = id,name\n2,german\n produced the single corrupted row {"id": 1, "name": "english2"} instead of two rows. The same happens for headless multipart CSVs.

The fix appends the missing newline to the last line of a CSV part when another part follows. Two deliberate boundaries:

• Non-tabular parts are untouched. For any other format the parts are chunks of one file (the loader's existing format != "csv" distinction), and inserting bytes would corrupt the reassembled file. A regression test pins this: b"\x00\x01NOEOL" + b"\x02\x03\n" still reassembles verbatim.
• The final part is left as-is — with no following part there is nothing to glue, so the stream stays byte-faithful there too.

Three tests added (header CSV, headless CSV, binary verbatim); the two CSV tests fail on main and pass with the fix. The full multipart scheme suite passes (8 passed, 4 skipped — the skips are the pre-existing remote/vcr ones). Also verified with CRLF endings and a three-part resource whose middle part lacks the newline.

While touching the constructor I switched its List[str] annotation to list[str] — ruff's UP006 flags the line once edited; the file's other pre-existing findings are left alone.