use map instead of broadcast for Tensor operators

[cvsvensson]

For deeply nested tensor products (4 or more), cheap operations like adjoint and size get slowdowns and start to allocate. The reason seems to be the combination of broadcasting and nested tuples. This can be fixed by replacing e.g. adjoint.(L.ops) with map(adjoint, L.ops). This PR replaces all such uses of broadcast with map for TensorProductOperator and TensorSumOperator.

Here are some benchmarks:

using SciMLOperators, BenchmarkTools
mo = MatrixOperator(rand(1, 1))
deep_prod = TensorProductOperator([mo for _ in 1:10]...)

@btime adjoint($deep_prod)
# before: 1.380 μs (29 allocations: 1.42 KiB)
# after : 12.713 ns (0 allocations: 0 bytes)

@btime $deep_prod + $deep_prod
# before: 7.850 μs (73 allocations: 3.08 KiB)
# after: 14.028 ns (0 allocations: 0 bytes)