Restore for variables and factors the module prefix

[dehann]

Restore for variables and factors the module prefix, e.g. "RoME.Pose2", or "Caesar.ScatterAlign". Fallback is to search for the fields in Main. FYI, this used to be in older DFG, probably just got lost in the shuffle.

BTW, fileDFG also does not have the module in factors, so also just Pose2Pose2. So, I would consider it a breaking change to to include it.

_Originally posted by @Affie i

[dehann]

From Affie:

Yes, that should definitely be there. If we run unique code without a package definition then we can either fall back to Main automatically, but first need to do a security audit on hte code. What if a user uploads a piece of malicious code and gets run through this evaluation. i think the security will be better by resitrcting to packages we know and trust as general rule, and then only deal with the Main exceptions.

Current deserialization calls eval which is very bad. I fixed it, but with new constraints I think we can live with.

[Affie]

Variables and beliefs (distributions) have the module prefix, it is only Factors that does not.
Easiest way forward is to only fix with the IIFTypes and RoMETypes upgrade. Deprecation can be handled by ignoring the module in old data for a while.

[Affie]

Hardcoding the "packed" prefix in factors is also not the best idea. We should probably use the actual type in the type field. That way a user does not have to use a specific name and a lot of the factors might not need a packed version (if <: AbstractPackedFactor doesn't cause issues).

[dehann]

[do IIFTypes, RoMETypes first]

Agree, that sounds like a good way to go.

[Packed name mangling]

I thought we are dropping the Packed prefix, and expressly calling user provided (un)pack for each factor only when going all the way to FactorCompute. When transporting FactorDFG objects then the actually mechnics of a factor are not unpacked. I understood these decisions means we can drop the "Packed" prefix, am I missing something?

[Affie]

As example,

• currently [de]serialization of a PackedPose2Pose2 use the fnctype field in FactorDFG .
• fnctype contains Pose2Pose2 and not PackedPose2Pose2 which makes deserialization impossible without adding the Packed prefix.

I'm saying [de]serialization should maybe use a type field in the packed observation (pending how JSON.jl v1 works)