fix(AutoCache): ignore normalisation in check

The normalisation constant is applied outside of the model, so if the
model is additive the normalisation parameter does not need to be
checked to see if the model needs re-evaluation. This makes a big
performance difference if fitting simultaneous datasets on the same with
bound parameters domain but with different normalisations.

src/meta-models/caching.jl

@@ -52,7 +52,7 @@ function _reinterpret_dual(
     reinterpret(DualType, view(v, 1:n_elems)), needs_resize
 end
 
-function invoke!(output, domain, model::AutoCache{M,T}) where {M,T}
+function invoke!(output, domain, model::AutoCache{M,T,K}) where {M,T,K}
     D = promote_type(eltype(domain), T)
 
     _new_params = parameter_tuple(model.model)
@@ -65,8 +65,14 @@ function invoke!(output, domain, model::AutoCache{M,T}) where {M,T}
 
     # if the parameter size has changed, need to rerun the model
     if (!out_resized) && (model.cache.size_of_element == sizeof(D)) && (same_domain)
+        # ignore the normalisation, since that's applied later
+        _pc, _np = @views if K === Additive
+            param_cache[2:end], _new_params[2:end]
+        else
+            param_cache, _new_params
+        end
         # if all parameters within some tolerance, then just return the cache
-        within_tolerance = all(zip(param_cache, _new_params)) do I
+        within_tolerance = all(zip(_pc, _np)) do I
             p, pm = I
             abs((get_value(p) - get_value(pm)) / p) < model.abstol
         end