Broadcast with missing destroys sparsity

[LilithHafner]

julia> f(x, y) = x
f (generic function with 1 method)

julia> f.(Diagonal([1,2,3]), nothing)
3×3 Diagonal{Int64, Vector{Int64}}:
 1  ⋅  ⋅
 ⋅  2  ⋅
 ⋅  ⋅  3

julia> f.(Diagonal([1,2,3]), 6)
3×3 Diagonal{Int64, Vector{Int64}}:
 1  ⋅  ⋅
 ⋅  2  ⋅
 ⋅  ⋅  3

julia> f.(Diagonal([1,2,3]), missing)
3×3 Matrix{Int64}:
 1  0  0
 0  2  0
 0  0  3

This is due to the use of missing as a sentinel in LinearAlgebra/src/struguredbroadcast.jl internals.