Fix stale free state handling in custom allocator for CUDA and HIP

[bwlang]

without this fix i was seeing GPU runs crashing with errors like these:

• "illegal memory access" errors inside RELION's GPU allocator (custom_allocator.cuh)
• segfaults in CudaCustomAllocator::_free
• an out-of-memory report asking for 564 TB
• "No orientation was found as better than any other"
• NVIDIA Xid 31 and Xid 13 ("Out Of Range Address") in the kernel log, on two different GPU

with this fix it's working.