Support empty varlen batches

[0z5a]

Summary

• Define zero-length sequence semantics for cu_seqlens with repeated offsets.
• Avoid constructing invalid zero-extent TMA descriptors for all-empty batches.
• Preserve initial_state in final_state, or write zeros when no initial state is supplied.

Validation

• RTX 5080 (SM120): python -m pytest tests/test_empty_varlen.py -vv — 32 passed.
• Before this change, mixed-empty batches passed but all-empty batches aborted in the C++ extension.

Stacked on MoonshotAI#29; update the base after MoonshotAI#29 merges.