don't use generated function implimentation of naivemul for sparse matrices

[oscardssmith]

I'm in general pretty skeptical of the naivemul introduced in #57, but this at least makes it dramatically better for sparse matrices. This improves, but does not fix SciML/OrdinaryDiffEq.jl#2121.

# before
julia> @time exponential!(sprand(500,500,.001), ExpMethodGeneric());
  9.914123 seconds (52 allocations: 13.983 MiB)

#after
@time exponential!(sprand(500,500,.001), ExpMethodGeneric());
  2.100566 seconds (260 allocations: 14.785 MiB)