[graph] Disallow input dependencies without registered update

[arntanguy]

This PR makes sure that we have a registered update function corresponding to an input dependency.
The reasoning is to avoid a potentially misleading edge case when inheriting from a function and declaring updates.

struct BaseFun : public Node
{
SET_UPDATES(BaseFun, Value)

BaseFun() {
  registerUpdates(Update::Value, &BaseFun::updateValue_);
  addInputDependency<BaseFun>(Update::Value, source, Output);
}
};

Currently it is possible to do

struct DerivedFun : public BaseFun
{
SET_UPDATES(DerivedFun, Value)

DerivedFun() {
  addInputDependency<BaseFun>(Update::Value, source, Output);
}
};

Here DerivedFun::Update::Value (=1) is different from BaseFun::Update::Value (=0), and only BaseFun::Update::Value has a registered update function. Yet it is possible to declare an input dependency for DerivedFun::Update::Value that does not have a corresponding update function. This will only fail at runtime if:

• The update is called => crash in release, false assert in debug
• The graph's Log is printed: there is no update corresponding to DerivedFun::Update::Value

Instead we should either decare

struct DerivedFun : public BaseFun
{
SET_UPDATES(DerivedFun, Value)

DerivedFun() {
  registerUpdates(Update::Value, &DerivedFun::updateValue_); // explicitely register the update for DerivedFun::Update::Value
  addInputDependency<BaseFun>(Update::Value, source, Output);
}
};

or

struct DerivedFun : public BaseFun
{
DerivedFun() {
  addInputDependency<BaseFun>(Update::Value, source, Output); // use update signal and function from parent class
}
};

This PR merely ensures that we cannot have a call to addInputDependency without a corresponding update function. It however forces the user to always call registerUpdates before addInputDependency, which in practice is already the case with only very few exceptions.