This PR ensures that the check-omp-pragmas check only checks LightGBM's own first-party source code, not:

• code pulled in via submodules


• other untracked code in the repo (e.g. code in Python virtual environments)

For some situations, this makes local runs of pre-commit run --all-files faster and removes a source of false positives.

While touching this, also added a few more C/C++ extension types to the check:

• 
  .cu (CUDA C++ source files)


• 
  .cuh (CUDA C++ headers)


• 
  .tpp (C++ template files)

Notes for Reviewers

How I found this

I've recently switched from using conda for Python development here to virtualenvs, and frequently do stuff like this:

python -m venv .venv

source .venv/bin/activate

pip install numpy scipy scikit-learn

# etc., etc.

When I started doing that, I found that check-omp-pragmas at time was noticably slower, and occasionally reported findings in files that weren't part of LightGBM.

How I tested this

Manually removed a num_threads() class from the following places:

• 
  include/LightGBM/bin.h here


• some places in external_libs/eigen
  


• a copy of bin.h placed in .venv/
  

Saw what I expected... only the first-party LightGBM one was reported and the hook failed.

$ pre-commit run --all-files check-omp-pragmas

check-omp-pragmas........................................................Failed

- hook id: check-omp-pragmas

- exit code: 1


checking that all OpenMP pragmas specify num_threads()

include/LightGBM/bin.h:68:  #pragma omp parallel for schedule(static)

Found '#pragma omp parallel' not using explicit num_threads() configuration. Fix those.

For details, see https://www.openmp.org/spec-html/5.0/openmpse14.html#x54-800002.6