Scalar multiplication may not preserve bandwidths

[jishnub]

julia> B = brand(Complex{Int8}, 3,3, 5,5)
3×3 BandedMatrix{Complex{Int8}} with bandwidths (5, 5):
 -57-25im     0-61im   18-6im
 -94-43im   102+18im  106-40im
 110+26im  -114+15im   75-50im

julia> C = 1*B
3×3 BandedMatrix{Complex{Int64}} with bandwidths (2, 2):
 -57-25im     0-61im   18-6im
 -94-43im   102+18im  106-40im
 110+26im  -114+15im   75-50im

While the allocation is lower if there are fewer bands, altering the bandwidths makes it harder to loop over the bands. If the user has specified empty bands, perhaps these should be preserved at the expense of allocations?

[dlfivefifty]

I feel like this is a fringe situation with trade-offs so would need a strong argument to change current behaviour.

[jishnub]

This should be consistent across eltypes, whichever direction we decide to proceed. Currently,

julia> B = brand(Int8, 3,3, 5,5);

julia> bandwidths(1B)
(5, 5)

julia> B = brand(Complex{Int8}, 3,3, 5,5);

julia> bandwidths(1B)
(2, 2)