more performant SparseMatricCSC constructors for structured matrices

[dkarrasch]

This is a follow-up on #32466, implementing what was suggested here #10843 (comment). I was experimenting with all kinds of high-level things, like repeat, and collect of ranges etc., but nothing was working as fast with as little allocations as simply filling undefined vectors of appropriate type (especially repeat was really bad!). Some benchmarks are below. The call to sparse corresponds to the current master code (i.e., "before PR"), the calls to SparseMatrixCSC call functions of this PR ("with PR"). For larger sizes the gain can be a factor larger than 10.

julia> using SparseArrays, LinearAlgebra, BenchmarkTools

julia> m = 1000
1000

julia> T =  Tridiagonal(randn(m-1),randn(m),randn(m-1));

julia> @btime sparse($T);
  27.698 μs (25 allocations: 188.83 KiB)

julia> @btime SparseMatrixCSC($T);
  3.931 μs (7 allocations: 55.09 KiB)

julia> T = SymTridiagonal(randn(m),randn(m-1));

julia> @btime sparse($T);
  27.504 μs (25 allocations: 188.83 KiB)

julia> @btime SparseMatrixCSC($T);
  3.880 μs (7 allocations: 55.09 KiB)

julia> B = Bidiagonal(rand(m), rand(m-1), :U);

julia> @btime sparse($B);
  21.705 μs (20 allocations: 134.38 KiB)

julia> @btime SparseMatrixCSC($B);
  2.664 μs (5 allocations: 39.56 KiB)

julia> B = Bidiagonal(rand(m), rand(m-1), :L);

julia> @btime sparse($B);
  21.365 μs (20 allocations: 134.38 KiB)

julia> @btime SparseMatrixCSC($B);
  2.632 μs (5 allocations: 39.56 KiB)