[BUG] RMM CI should not only test PTDS mode

[harrism]

Recently a decision was made to build and test cuDF with Per-Thread Default Stream set to ON, because this is a more conservative test than OFF and therefore it should be sufficient to test cuDF only with PTDS ON.

However, for some reason that change was also applied to RMM which is NOT a valid testing approach since unlike cuDF, RMM has separate code paths for PTDS=ON and PTDS=OFF. If we only test with PTDS ON, then we will not be testing both paths.

Instead, librmm includes CMake code to compile some of its tests in both modes, creating separate PTDS and non-PTDS test binaries. So for C++ / gtesting, PTDS should not be globally configured for RMM.

For Python testing, we need to change to building the Cython bindings twice (once with PTDS=ON and once with PTDS=OFF) and run the pytests twice.

[harrism]

Fixed by #603