[maleadt]

Sparse mapreduce had two bugs, both reproducible with JLArrays:

• the row and column kernels started from op(init, init) with init = zero(eltype), so maximum of a row whose stored entries are all negative returned 0 (also for the full reduction, which reduced the per-row results with the same zero init);
• implicit zeros were folded in as nzeros * f(0), which is only right for +: mapreduce(x -> x + 1, max, A) for a 3×3 matrix with two stored entries gave 3.0 instead of 1.0.

Each slice now reduces its stored values starting from the first one and folds in its implicit zeros afterwards with op itself (by repeated squaring, or once for idempotent operators such as max), which is what SparseArrays does. The full reduction reduces the stored values with the dense mapreduce and folds the zeros in on the host. An explicit init is combined with each result, empty slices follow Base (init, or the reduction of an empty collection), and dims=2 returns an m×1 matrix like Base instead of a vector.

Tests: all-negative matrices with and without implicit zeros (maximum/minimum over every dims), non-zero-preserving functions with *, max and min, init, an operator without a known neutral element, result shapes, and empty dimensions.

Stacked on #795, below the sparse redesign in #745; it can go into 11.x on its own.