possible issue with u64->f64 conversion in rand/float.rs

[jasondark]

The function into_float_with_exponent implements the standard bit-pattern hack to grab the 53 significant bits and subtract 1. However, according to http://xoshiro.di.unimi.it/ (see section "Generating uniform doubles in the unit interval") this halves the range of possible values by always resulting a least significant bit of zero. It is proposed to just use the more straight-forward (myint>>11) as f64 / (1u64 << 53) as f64 instead.

[pitdicker]

@jasondark Thank you for opening this issue. Just a couple of days late for the party, see #416 which had similar concerns.

In a way both methods are imperfect, as floats do not have a fixed precision. But we changed the default method to the traditional multiply approach in #420.