Dataclasses for Recogniser, Ranker and Linker output data types

[thobson88]

AIM: replace nested dictionaries with appropriate data structures implemented as Python dataclasses.

Subtask of #276.

[thobson88]

Dataclass structure

Key:

• Fields (attributes) above the line
• Methods (functions) below the line
• Arrows denote subclasses
• Diamonds denote composition
• Green indicates an abstract base class
• Orange indicates a concrete class

[thobson88]

Some advantages of dataclasses over nested dictionaries:

• code is more comprehensible, easier to discuss & reason about (by refering to classes by name)
• less error prone: explicit references to named data fields, rather than arbitrary indexing via magic strings/numbers.
• data provenance: dataclass instances retain all of the info and logic from which a prediction is generated, rather than just the final scores.
• algorithm clarity: e.g. all business logic lives in the appropriate Recogniser/Ranker/Linker subclasses and bookkeeping is done in the dataclasses.
• no need for utility functions to switch format between different dictionary representations
• automatic field type checking with pydantic.dataclasses
• automatic sorting in lists
• immutability
• pretty printing

[thobson88]

TODO: add fields latlon and wkdt_class into the WikidataLink class structure.