Use np.float64 to fix issue reading cached LD files

[jdblischak]

I confirmed with a local example that I was able to read a cached LD file that was created using np.float64. I also observed the following compared to the current np.float32 version:

• It required more memory to create the LD matrix
• The .npz file almost doubled in size
• The .gz file was unchanged (confirmed with zcat *.gz | md5sum)
• The checksum for the results file was also different compared to np.float32. There were minor differences in PIP, BETA_MEAN, and BETA_SD
• The results obtained from the cached LD file were identical to the initial np.float64 run

Close #63

[jdblischak]

Also, the tests in python test_polyfun.py passed