eltype deduction regression in Julia 1.3 and above for generic matmul

Julia 1.0.5:
```julia
julia> Real[1 2] * Real[1.5; 2.0]
1-element Array{Real,1}:
5.5
```

Julia 1.3.1:
```julia
julia> Real[1 2] * Real[1.5; 2.0]
1-element Array{Any,1}:
5.5
```

For reference, broadcasting deduces the optimal eltype:

```julia
julia> Real[1 2] .* Real[1.5; 2.0]
2×2 Array{Float64,2}:
1.5 3.0
2.0 4.0
```

Maybe generic mul should do the same, such that subsequent dispatches can be fast.