feat: add DM command safety layer

[aviralmittal7150]

Summary

• add configurable DM stroke limits and park thresholds to both configuration paths
• apply saturation/parking exactly once to final map and closed-loop commands, with status propagation through streaming and extended C APIs
• reject non-finite commands/configuration safely and clear outputs on mapping failures
• bump the public ABI version to 3.0.0 because the public config struct grew
• add Python bindings, documentation, CPU/CUDA safety tests, and a CUDA host-fallback map that uses the CPU safety authority

Issues fixed

• stream results previously reported OK after map safety had already clamped or parked commands
• closed-loop updates could be limited before gain/accumulation and early convergence could return unsafe pre-existing commands
• NaN/Inf command values bypassed the safety layer
• CUDA clamped an uninitialized DM buffer and had no parking/status parity
• rippra_dm_park was not reliably exported from the Windows shared library
• the Python binding had no status-bearing saturation/map interface or ABI guard

Verification

• clang -std=c11 -Wall -Wextra -Werror -I rippra/include -fsyntax-only on changed C sources/tests
• focused shared-library C safety test: 43/43 passed
• existing closed-loop C test: 10/10 passed
• python3 -m py_compile rippra/bindings/rippra.py rippra/tests/test_dm_safety_bindings.py
• python3 -m unittest rippra.tests.test_dm_safety_bindings (8 skipped because NumPy is not installed in this environment)
• git diff --check

Environment limitation

CMake and nvcc are not installed in this workspace, so CMake/CUDA builds could not be executed locally. The CUDA test is registered only when CUDA is available and skips with return code 77 when no GPU is present.