Improve FP64 accuracy

Expected outcome: All fp64 operations are bit-accurate as defined by PTX spec (https://docs.nvidia.com/cuda/parallel-thread-execution/index.html)

Comments:

• fp32 and lower is already accurate
• fp64 is pretty much only used in HPC, so there's been no pressing need to improve it. With ZLUDA not supporting sever gpus this has been somewhat low priority, but since mi450x is shaping up to be supported by ZLUDA, it's getting more relevant
• We have a test suite here to test it: https://github.com/vosen/ptx_tests