Make getindex for Bidiagonal O(1) in the worst-case

[mcognetta]

This is similar to #32328, but I separated them since that is a bug fix and this is just an edge case optimization.

In the worst case, getindex for Bidiagonal matrices is O(n). This PR makes it O(1).

The worst case is a Bidiagonal where the off diagonal is zero. In the worst case, iszero (which is O(n)) is called twice (once in istriu and once in istril).

This only really comes into play for large matrices, but this pr doesn't slow down getindex for small matrices, so it seems like a net win.

---

Here are some timings:

Master:

julia> B = Bidiagonal(rand(100000), zeros(99999), 'U')

julia> @time B[100000, 99999]
  0.000379 seconds (5 allocations: 176 bytes)
0.0

julia> @time B[99999, 100000]
  0.000008 seconds (5 allocations: 176 bytes)
0.0

This PR:

julia> B = Bidiagonal(rand(100000), zeros(99999), 'U')

julia> @time B[99999, 100000]
  0.000003 seconds (5 allocations: 176 bytes)
0.0

julia> @time B[100000, 99999]
  0.000007 seconds (5 allocations: 176 bytes)
0.0

[mcognetta]

CI Failure seem unrelated to this PR. Same with #32328

[dkarrasch]

Yes, this is how it should be.

[mcognetta]

Bump (and #32328)