Mask final word in map_open_close

[SRO-SA]

Fixes #6

Summary:

• Updates map_open_close to mask invalid tail bytes in the final partial word.
• Avoids the byte-by-byte fallback path and keeps the existing 32-bit word read path.
• Relies on the padded open/close bitmap allocation being zero-filled before the kernel runs.

Testing:

• I could not run CUDA or compute-sanitizer locally because my laptop does not have CUDA access.
• I reviewed the affected JSON Lines code path manually.
• Please rerun the initcheck reproducer from the issue to confirm the uninitialized read is resolved.