Dependent Parameters do not update the calculator

[rozyczko]

When a dependent Parameter is bound to a calculator through _callback (in InterfaceFactoryTemplate.generate_bindings), changes to
its independent component never propagate to the calculator, and the recomputed value is overwritten with the old calculator value on the next .value read.

The result: every constraint with a bound dependent parameter is completely neglected during fitting and simulation.

In easyreflectometry this breaks:

• MaterialMixture: fitting fraction does not change its value, and fraction has zero effect on the computed reflectivity (_sld/_isld are dependent parameters bound to the calculator),
• gradient-layer thickness/roughness constraints (each sub-layer's thickness/roughness is dependent on the assembly parameter),
• surfactant/bilayer area-per-molecule, solvation and roughness constraints.

No error or warning is raised anywhere.

[rozyczko]

The solution is to just call fset in _update, like we do in the value setter.

[rozyczko]

_update never calls the value setter - the setter explicitly raises for dependent parameters "This is a dependent parameter, its value cannot be set directly."). We are only allowed to change a dependent parameter by changing its independent co-parameter.
_update() writes straight into the internal scipp scalar

[rozyczko]

to reproduce:

from easyscience import Parameter

driver = Parameter(name='fraction', value=0.5, min=0.0, max=1.0)

dependent = Parameter.from_dependency(
    name='sld',
    dependency_expression='2 * a',
    dependency_map={'a': driver},
)

# Simulate InterfaceFactoryTemplate.generate_bindings: bind to external storage
calculator_storage = {'sld': dependent.value}  # 1.0
dependent._callback = property(
    fget=lambda: calculator_storage['sld'],
    fset=lambda v: calculator_storage.__setitem__('sld', v),
)

driver.value = 0.9  # triggers dependent._update()

print(dependent._scalar.value)    # 1.8  <- dependency recomputed correctly
print(calculator_storage['sld'])  # 1.0  <- BUG 1: calculator never updated (expected 1.8)
print(dependent.value)            # 1.0  <- BUG 2: getter restores stale value (expected 1.8)
print(dependent._scalar.value)    # 1.0  <- correct result has now been erased

[damskii9992]

Won't calling fset in the _update trigger new notify_observers causing an infinite loop?

[damskii9992]

The more I am looking at this, the more I want to make the proper Calculator Factory as discussed in #7 and #160. I REALLY don't wan't to use this legacy crap for easyimaging . . . .
Maybe I'll start writing it when I need it for easyimaging 🤔

[rozyczko]

Won't calling fset in the _update trigger new notify_observers causing an infinite loop?

Nope. There's no loop

1. fset is not the Paramter's value setter. Look at interface_factory.py make_prop. fset is the calculator's update function, defined entirely on the library side.
2. _notify_observers calls _update on each observer, that is parameter dependent on this one, not the ones it depends on.
   for b=a*2, if a changes, only b.update() is called, which updates other dependencies of b, say c=b+1

I guess the only potential hole is if we have a library setter_fn which in turn writes back into parameter.value explicitly... But this issue is valid even before my fix....