[BUG] get_primary_cuda_context static map is unsafe in three ways

[wence-]

1. The CudaPrimaryContext object has a destructor which calls into the driver (cuDevicePrimaryCtxRelease). But since these objects are stored in a static map, the destructor will run below main, and this is UB per the cuda programming guide
2. get_primary_cuda_context is not thread-safe since access to the static map is not protected by a lock.
3. (This came out of reviewing Add resource_ref versions of get/set_current_device_resource rmm#1598)

Since kvikio is header-only, we rely on the linker to disambiguate inline functions that have (or return) static references. To do this, the relevant function must have __attribute__((visibility("default"))). If not, then if kvikio is used in two different DSOs, the function will appear twice, and there will be two static objects.

By default, I believe all rapids libraries are compiled with -fvisiblity-inline-hidden. So kvikio::get_primary_cuda_context (which is inline) will be __attribute((visibility("hidden"))). Thus, if kvikio is used from two separate DSOs, calls from the respective DSOs to get_primary_cuda_context will reference different global statics (since the functions are not the same).

To solve these:

1. We just need to leak the contexts, I think.
2. We can add a lock
3. We should explicitly mark this function as __attribute__((visibility("default")))