Align softmax_online scaffold with issue #38

[debashishc]

Summary

• switch softmax_online to backend registry flow (ref + kernel) with simpler dispatch
• add kernel module from PR26 as explicit non-production reference (with credit header)
• enforce strict dim=-1 for softmax_online op/ref paths
• exercise kernel backend in tests (forward + backward coverage)
• align softmax benchmarks with backend selection APIs instead of env mode

Validation

• uv run ruff check forge_cute_py/ops/softmax_online.py forge_cute_py/ref/softmax_online.py tests/test_softmax_online.py
• uv run pytest tests/test_softmax_online.py -q
• uv run pytest -q