fix: don't split session JSONL on Unicode line separators

[rian-dolphin]

Session files are JSONL, but the reader split them with str.splitlines(), which also breaks on U+2028/U+2029/U+0085. These are characters pydantic writes unescaped into JSON strings. One message containing such a character (e.g. from a file read into a tool result) splits a valid line into two malformed fragments, so every subsequent session load/resume raises SessionJsonlError — the session is permanently bricked. Fix: split on \n only, which also retroactively repairs already-broken files; same bug fixed in the session index reader.

To reproduce, run this snippet on main (fails with SessionJsonlError) and on this branch (prints one entry). Note the append succeeds either way — the corruption is only visible on the next read, and typing/pasting the character into the TUI won't trigger it because the input path sanitizes it; the script drives the storage API directly, as a tool result would:

uv run python - <<'EOF'
import asyncio, tempfile
from pathlib import Path
from tau_agent.messages import UserMessage
from tau_agent.session.entries import MessageEntry
from tau_agent.session.storage import JsonlSessionStorage

content = "Here is the error:" + chr(0x2028) + "TypeError: cannot read properties of undefined"

async def main():
    with tempfile.TemporaryDirectory() as tmp:
        storage = JsonlSessionStorage(Path(tmp) / "session.jsonl")
        await storage.append(MessageEntry(message=UserMessage(content=content)))
        print("append: OK (write never fails)")
        entries = await storage.read_all()  # raises SessionJsonlError before this fix
        print(f"read_all: OK, {len(entries)} entry")

asyncio.run(main())
EOF

[alejandro-ao]

looks great. thanks!!