Spectral element operator patterns (possible improvements)

Following a review comment by [charleskawczynski](https://github.com/charleskawczynski), I'll leave this suggestion in an issue for future visibility - we currently use `@simd` in the spectral operator blocks, but an alternative is as follows:

###

```julia
e1 = length(input) ≥ 1 ? (mapreduce(k -> D[i, k] ⊠ input[1][k, j, vt], ⊞, 1:Nq), mapreduce(k -> D[j, k] ⊠ input[1][i, k, vt], ⊞, 1:Nq)) : ()
e2 = length(input) ≥ 2 ? (mapreduce(k -> D[i, k] ⊠ input[2][k, j, vt], ⊞, 1:Nq), mapreduce(k -> D[j, k] ⊠ input[2][i, k, vt], ⊞, 1:Nq)) : ()
e3 = length(input) ≥ 3 ? (mapreduce(k -> D[i, k] ⊠ input[3][k, j, vt], ⊞, 1:Nq), mapreduce(k -> D[j, k] ⊠ input[3][i, k, vt], ⊞, 1:Nq)) : ()
t = (e1...,e2...,e3...)

if length(input) == 1 # check types
return Geometry.Covariant12Vector(t...)
elseif length(input) == 2
return Geometry.AxisTensor((Geometry.Covariant12Axis(), Geometry.UVAxis()), t)
else
return Geometry.AxisTensor((Geometry.Covariant12Axis(), Geometry.UVWAxis()), t)
end
```

_Originally posted by @charleskawczynski in https://github.com/CliMA/ClimaCore.jl/pull/1736#discussion_r1710014985_

The issue names only the spectral operator blocks and the current @simd approach; start by locating those blocks and reviewing the linked discussion in PR 1736. Compare the proposed mapreduce pattern for correctness and performance, and consider the work done only after its trade-offs are established and validated.

Written by the indexing model from the issue text.