Avoid overflow in float interval medians

[zayenz]

An unbounded FlatZinc float variable can enter an all-negative interval whose median overflows to negative infinity. The split brancher then has one failing alternative and one alternative that leaves the domain unchanged, so search repeats indefinitely.

Compute finite medians with the existing nearest-rounding policy, scaling the operands only near the numerical limits to avoid overflow. Semi-infinite intervals retain their legacy extended endpoint, the fully unbounded interval has symmetric median zero, and singleton infinities remain unchanged.

Regression coverage includes positive, negative, and mixed intervals at the Float limits; the one-ULP closest-representation case; semi-infinite and infinite intervals; signed zero; and subnormal intervals. The reported model goes from 977,266 nodes without a solution after one second to 54 nodes and a solution in under one millisecond.

Checks:

• Float::Basic (20 iterations)
• all Float::Branch tests (20 iterations)
• original MiniZinc model compiled with MiniZinc 2.9.7 and run through current fzn-gecode (1 solution, 54 nodes)
• 500,000 randomized finite interval pairs against an MPFR exact-midpoint oracle under all four IEEE rounding modes
• git diff --check

Closes #164.