NSNumber: hash by numeric value rather than truncating

[DTW-Thalion]

-[NSNumber hash] cast doubleValue to an integer, so every value sharing an integral part collided: 1.5 and 1.7 hashed alike.

This reduces doubleValue modulo a Mersenne prime, after the manner of CPython, so the fractional part contributes and an integral value still hashes to that integer.

The hash stays a function of doubleValue because that is what -compare: and -isEqual: use. NSNumber equality promotes to double, so a value held as int, float or double compares equal across those types and must hash equally. Hashing the raw storage, as #535 first tried, breaks that: the bit patterns of an int and a double of the same value differ, so equal numbers get different hashes and fail as dictionary keys.

One consequence worth noting: equality is not transitive across the precision boundary. A long long above 2^53 compares equal to the double it rounds to, so its hash must be that of the double, and distinct large integers can collide. That follows -compare: and cannot be avoided without changing comparison.

Tests/base/NSNumber/NSNumber_hash.m covers the -hash/-isEqual: contract, fractional scattering and non-finite values. 862 tests pass under both the GNU and libobjc2 runtimes.

This is an option for #535, following the doubleValue-based approach discussed there.

[rfm]

Great, Given that this is the approach I found looked best when I searched for alternative hashing solutions, I have nothing to complain about here. I know Hugo was going to come back after reviewing the literature, but this is a good solution (and if he finds a better one we can always change later).