A much better hash function would be to just reinterpret the bit pattern in the float as an integer, I think. Maybe xor'ing the high and low ends if the floats are longer than INT32 (which they typically are on 64 bit architectures).
Don't know if it could be a compat problem changing it though. Strictly speaking it shouldn't be, considering the doc for hash_value, but that's not necessarily the end of the story.
