[BUG] Add in float, double, timestamp, and date support to murmur3

[revans2]

I am filing this as a bug because it is technically a latent bug/feature regression that was found as a part of #1910
With better testing we found that double and float types did not do the correct thing for either NaN values or for -0.0. The NaN values have to be normalized to a single NaN but the -0.0 should not be. In #1910 I updated the murmur3 code to normalize the NaN values, but the hash code itself for what ever reason was treating the -0.0 as 0.0 and outputting the same hash for both. The Spark code extracts the bit representation of these values as either a long or an int and then hashes those. In cudf 0.5 this is now possible with bitCast and we should use it to hopefully get around the issues with float and double. Because hash partitioning now also depends on murmur3 it and its tests also need to be updated.