Preserve C++ template arguments in NCU kernel matches

[yoshifuminakamura]

Summary

• parse demangled CUDA kernel names with template- and argument-list awareness
• keep C++ template arguments in generated NCU regex matches while dropping return types and top-level function arguments
• add a fixture for distinct templated kernels with parenthesized non-type template arguments

Tests

• python3 -m py_compile scripts/profiling/generate_ncu_plan.py
• bash scripts/tests/test_ncu_plan_generation.sh
• shellcheck -S error scripts/tests/test_ncu_plan_generation.sh
• bash scripts/tests/run_profile_data_shell_tests.sh