SUGGESTION: Run the OpenCL tests under Oclgrind in CI to catch data races and invalid memory accesses in kernels #3430

Description

Motivation. CI runs the OpenCL tests on one GPU. A kernel bug that this device happens to tolerate stays invisible. Example: #3429, where the cholesky_decompose kernel returns a wrong factor on an AMD GPU because its barriers fence local memory only.

Oclgrind is an OpenCL device simulator. It needs no GPU and checks what real devices do not:

• data races between work items, including a barrier with the wrong fence flag (--data-races),
• reads and writes outside a buffer (always on),
• use of uninitialized values (--uninitialized).

It runs our test binaries unchanged:

oclgrind --data-races --log races.log ./test/unit/math/opencl/cholesky_decompose_test

On develop this reports read-write races in cholesky_decompose, already for the 3×3 case, and nothing once the barriers are fixed. A first local pass also reports races in tridiagonalization_householder.

Suggested steps.

1. Add oclgrind to the CI image.
2. Add a CPU-only stage that builds test/unit/math/opencl with STAN_OPENCL=true and runs each test binary through oclgrind --data-races --log <file>. runTests.py already prefixes MPI tests with mpirun; the same mechanism can prefix oclgrind.
3. Fail the stage when a log is not empty.
4. Start it as non-blocking and make it required once the existing reports are fixed.