Add challenge 94: SSM Selective Scan (Medium)

[claude]

Summary

• Adds challenge 94: SSM Selective Scan (Medium difficulty), the core recurrence at the heart of Mamba-style state space models
• Solvers implement the full selective scan: discretize A using exp(delta * A), update a vector-valued hidden state h of size d_state for each of d_model independent channels, and compute output via the C projection plus a skip connection
• Teaches parallel scan algorithms (the naive sequential loop is O(seq_len); a good GPU solution uses a parallel prefix scan), multi-dimensional hidden state management, and memory layout for 4D tensors
• Clearly distinct from challenge 82 (Linear Recurrence), which has a scalar hidden state — here the state is d_state-dimensional per channel, requiring a C-projection output step
• All 6 starter files provided; CUDA reference solution validated against the live T4 platform (✓ All tests passed)

Test plan

• pre-commit run --all-files passes (black, isort, flake8, clang-format, mojo format)
• Example test passes on T4 via run_challenge.py --action run
• All functional + performance tests pass on T4 via run_challenge.py --action submit
• Checklist in CLAUDE.md verified: HTML starts with <p>, correct <h2> sections, <pre> example, performance bullet, SVG diagram, 10 functional test cases, performance test fits 5× in 16 GB VRAM

🤖 Generated with Claude Code