fix perf issue loadmodel!

[CarloLucibello]

Fix performance issue mentioned in #2239 (comment)

When loading from a flux model into another everything is fine:

using Metalhead, Flux

m1 = ResNet(18)
m2 = ResNet(18)
@time Flux.loadmodel!(m2, m1) # warmup
@time Flux.loadmodel!(m2, m1)
#  0.003388 seconds (23.01 k allocations: 2.157 MiB) # fast on both master and this PR

When loading from a similar struct such as nested named tuples instead (as proposed in #2239) perf on master is very bad:

using Functors

function state(x)
    if Functors.isleaf(x)
        return x
    else
        return map(state, Functors.children(x))
    end
end

s = state(m1)
@time Flux.loadmodel!(m2, s) # warmup
@time Flux.loadmodel!(m2, s)
# 69.507712 seconds (1.06 G allocations: 40.744 GiB, 4.74% gc time) # master
# 0.002409 seconds (6.85 k allocations: 1.109 MiB)                               # this PR

The perf problem was due to huge strings being created at each recursion step as part of error messages.

@theabhirath

Also removed the filter argument, which was not documented nor used in practice.

[theabhirath]

filter is important for Metalhead. The motivation is described in #2041

[CarloLucibello]

I'll revert the filter removal. Can we make it documented API?

[darsnack]

I don't have an issue with that, but I recall that last time there was some bikeshedding and we opted to keep it private. If it's still an issue, no need to let it hold up this PR.

done