hashCode for floating points has a very bad distribution

[gzm0]

We cannot implement hashCode for java.lang.{Float, Double} the way Java specifies, because these boxes equal the integral-value boxes if they itself contain the corresponding integer value.

Therefore, we must have:

new Double(4.0).hashCode == new Integer(4).hashCode

Currently we use _.toInt as hashCode but that arguably gives very bad distribution. We should consider using the implementations in BoxesRunTime.

[sjrd]

The implementations of BoxesRunTime isn't going to help, since it relies itself on Double.hashCode. Basically what it does is: 1) is it a Double with an Int value -> use Integer.hashCode. 2) Otherwise, use Double.hashCode.

We could do something like that, of course. But still, what should we use for the Double case? We cannot use doubleToLongBits like the JVM specifies because this method is not implementable (without typed arrays).

[gzm0]

This is now just dependent on #1284