Merging of ComponentArrays

Thanks for the really useful package! It has helped a lot to clean up our model code.

One functionality I missed is to merge component arrays. With tuples we can do:

t1 = (a=1, b=2, c=3)
t2 = (a=111, d=444)
merge(t1, t2)            # (a = 111, b = 2, c = 3, d = 444)

I can simulate this behavior, but I think my implementation is not very optimal:

function merge(a::T, b::T) where T <: ComponentVector
    ComponentVector(merge(NamedTuple(a), NamedTuple(b)))
end
function merge(a::ComponentVector, b::NamedTuple) 
    ComponentVector(merge(NamedTuple(a), b))
end
function merge(a::NamedTuple, b::ComponentVector) 
    ComponentVector(merge(a, NamedTuple(b)))
end

ca1 = ComponentVector(a=1, b=2, c=3)
ca2 = ComponentVector(b=22, d=44)

merge(ca1, ca2)     # ComponentVector{Int64}(a = 1, b = 22, c = 3, d = 44)
# also useful to add a parameter
merge(ca1, (;new=222)) # ComponentVector{Int64}(a = 1, b = 2, c = 3, new = 222)

# it works with ForwardDiff but with with Zygote
f(x) = sum(merge2(ca1, x))
f(ca2)
Zygote.gradient(f, ca2)

A good use case would be optimizing some parameters while keeping others fix:

foo(ca) = ca.a + ca.b + ca.c + ca.d
ca_fix = ComponentVector(a=1, b=2)
# optimize only parameter 'c' and 'd'
optim(ca_opt -> foo(merge(ca_fix, ca_opt))
      ...
    )

Start by reviewing the ComponentVector and Julia merge behavior, then trace how ComponentArrays are constructed from NamedTuples. Use the examples in the issue to define coverage for merging two ComponentVectors and combinations with NamedTuples, including the optimization and differentiation cases. Done means the requested merges work without breaking the shown ForwardDiff and Zygote use cases.