Eliminate double-promotion warnings.

Clang currently complains about implicit conversions, e.g.
```
test/packetmath.cpp:680:59: warning: implicit conversion increases floating-point precision: 'typename Eigen::internal::random_retval<typename Eigen::internal::global_math_functions_filtering_base<double>::type>::type' (aka 'double') to 'long double' [-Wdouble-promotion]
          data1[0] = Scalar((2 * k + k1) * EIGEN_PI / 2 * internal::random<double>(0.8, 1.2));
                                                        ~ ^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
test/packetmath.cpp:681:40: warning: implicit conversion increases floating-point precision: 'float' to 'long double' [-Wdouble-promotion]
          data1[1] = Scalar((2 * k + 2 + k1) * EIGEN_PI / 2 * internal::random<double>(0.8, 1.2));
```

Modified to explicitly cast to double.