Improve type inference for adjoint/transpose of nested matrices

[jishnub]

This PR special-cases the return-types of the adjoint and transpose for SMatrix and MMatrix. This isn't ideal, but currently Julia doesn't infer the return types of certain recursive calls with changing types, and this PR helps with type-inference in such cases.

An example (using this PR):

julia> S = SMatrix{3,2,SMatrix{1,2,Int,2},6}(fill(SMatrix{1,2,Int,2}(1:2), 6))
3×2 SMatrix{3, 2, SMatrix{1, 2, Int64, 2}, 6} with indices SOneTo(3)×SOneTo(2):
 [1 2]  [1 2]
 [1 2]  [1 2]
 [1 2]  [1 2]

julia> @inferred adjoint(S)
2×3 SMatrix{2, 3, SMatrix{2, 1, Int64, 2}, 6} with indices SOneTo(2)×SOneTo(3):
 [1; 2;;]  [1; 2;;]  [1; 2;;]
 [1; 2;;]  [1; 2;;]  [1; 2;;]

Whereas, on master,

julia> @inferred adjoint(S)
ERROR: return type SMatrix{2, 3, SMatrix{2, 1, Int64, 2}, 6} does not match inferred return type SMatrix{2, 3, _A, 6} where _A