Specialize isbanded for StridedMatrix

[jishnub]

This improves performance, as the loops in istriu and istril may be fused to improve cache-locality.
This also changes the quick-return behavior, and only returns after the check over all the upper or lower bands for a column is complete.

julia> using LinearAlgebra

julia> A = zeros(2, 100);

julia> @btime isdiag($A);
  340.626 ns (0 allocations: 0 bytes) # nightly v"1.12.0-DEV.1571"
  523.105 ns (0 allocations: 0 bytes) # this PR

julia> A = zeros(100, 2);
  340.626 ns (0 allocations: 0 bytes)  # nightly 
  39.155 ns (0 allocations: 0 bytes) # this PR

julia> @btime isdiag($A);

julia> A = zeros(100, 100);

julia> @btime isdiag($A);
  6.616 μs (0 allocations: 0 bytes) # nightly
  2.932 μs (0 allocations: 0 bytes) # this PR

julia> A = diagm(0=>1:100);

julia> A[3,4] = 1;

julia> @btime isdiag($A);
  2.759 μs (0 allocations: 0 bytes) # nightly
  196.156 ns (0 allocations: 0 bytes) # this PR

The fact that the fat-matrix case is slower is unfortunate, and I'm unsure why it's much slower. Perhaps we need a cutoff here to call the fallback implementation?
Edit: the performance issue would be resolved by #56492.

A similar change is added to istriu/istril as well, so that

julia> A = zeros(2, 100);

julia> @btime istriu($A);
  8.083 ns (0 allocations: 0 bytes) # nightly
  9.307 ns (0 allocations: 0 bytes) # this PR

julia> @btime istril($A);
  377.766 ns (0 allocations: 0 bytes) # nightly
  474.221 ns (0 allocations: 0 bytes) # this PR

julia> A = zeros(100, 2);

julia> @btime istriu($A);
  146.085 ns (0 allocations: 0 bytes) # nightly
  38.764 ns (0 allocations: 0 bytes) # this PR

julia> @btime istril($A);
  8.577 ns (0 allocations: 0 bytes) # nightly
  10.888 ns (0 allocations: 0 bytes) # this PR

julia> A = zeros(100, 100);

julia> @btime istriu($A);
  3.290 μs (0 allocations: 0 bytes) # nightly
  1.677 μs (0 allocations: 0 bytes) # this PR

julia> @btime istril($A);
  3.353 μs (0 allocations: 0 bytes) # nightly
  1.661 μs (0 allocations: 0 bytes) # this PR