Split the `invalidate` method in two phases: in `invalidateCheck` it
just checks if the model diff warrants the execution of the actual
`invalidate` method, which requires fetching the custom invalidation
data from storage. The actual `invalidation` method remains the same.
This should make invalidation faster for storage providers that store
the invalidation remotely.
Move out of the `runPipe`/`runAnalysis` functions the logic that
extracts the container from the vector type. This logic is now
delegated to `containerVectorToTuple` which converts the vector type to
a tuple of references. This allows python's `run{Pipe,Analysis}` to
safely call `nanobind::gil_scoped_release` which releases the GIL for
the entire duration of the `runPipe`/`runAnalysis` execution.
Add an enum to explicitly specify the `TaskArgumentAccess` of a given
`Pipe`. This is needed because in some cases it's required to manually
specify the `READ` or `WRITE` access.
Rework the `AnalysisRunner` and `PipeRunner` structs into single
functions, these do not need to have a template parameter specifying how
to extract containers as that logic has been delegated to a newly
introduced trait called `ExtractContainerFromList`.
Add the needed infrastructures that allow ephemeral classes (`PipeRun`s)
to be wrapped as a regular pypeline `Pipe` class. These automatically
have their dependencies tracked and propagated to the pypeline.
Giacomo Vercesi