Follow-up of #8442.

In #8442 we left room like this and this to also support using managed memory for UMP. However, there are a few minor kinks we need to pan out first:

• We need another set of numpy allocator APIs backed by managed memory
• We need to allow malloc_managed to be used when UMP is enabled
• We need to allow, at test start/collection time, setting managed memory instead of system memory
• And of course, we need to confirm on G+H locally (since currently it's not accessible through the CI) that all tests pass