Supporting strided ranges

[fuchsto]

From issue #3:

Consider the following use cases:

// assuming row-major storage
dash::Matrix<double, 2> matrix(...);

auto lcol = matrix.local.col(0);
auto lrow = matrix.local.row(0);
auto n_lcol = std::distance(lcol.begin(), lcol.end());
auto n_lrow = std::distance(lcol.begin(), lrow.end());

std::vector<double> copy_lrow(n_lrow);
std::vector<double> copy_lcol(n_lcol);

// works:
dash::copy(lrow.begin(), lrow.end(), copy_lrow.data());
// fails:
dash::copy(lcol.begin(), lcol.end(), copy_lcol.data());

Slicing columns from a matrix with row-major storage order is fine, but copying this slice is not.
This is a missing feature: DASH does not support strided copying yet.

[ 0 1 2 3 ]  <- row contiguous in memory   | 8 | 4 | 0 |  <- column is strided, requires packing
[ 4 5 6 7 ]                                | 9 | 5 | 1 |     into temporary buffer [ 0 4 8 ] or
[ 8 9 0 1 ]        -- rot 90° -->          | 0 | 6 | 2 |     single copy operations for every
                                           | 1 | 7 | 3 |     value

This affects the specification of at least the following interfaces:

• Implementations of dash::copy should support strided input ranges, the interface must not change
• dash::local_range should be extended by dash::local_ranges which resolves a list of local ranges in
  a global input range
• dash::local_range should return the first local range in a global input range