Add barebones Artifact abstraction

Adds the base interface without supplying a default implementation.
After it is checked in, we can think about whether we want to add a base
implementation, e.g. via fsspec.

src/nnbench/core.py

@@ -2,26 +2,51 @@
 
 from __future__ import annotations
 
+import os
 from dataclasses import dataclass, field
-from typing import Any, Callable, Iterable
+from typing import Any, Callable, Generic, Iterable, TypeVar
+
+T = TypeVar("T")
 
 
 def NoOp(**kwargs: Any) -> None:
     pass
 
 
-# TODO: Should this be frozen (since the setUp and tearDown hooks are empty returns)?
-@dataclass(init=False)
-class Params:
+class Artifact(Generic[T]):
     """
-    A dataclass designed to hold benchmark parameters. This class is not functional
-    on its own, and needs to be subclassed according to your benchmarking workloads.
+    A base artifact class for loading (materializing) artifacts from disk or from remote storage.
 
-    The main advantage over passing parameters as a dictionary is, of course,
-    static analysis and type safety for your benchmarking code.
+    This is a helper to convey which kind of type gets loaded for a benchmark in a type-safe way.
+    It is most useful when running models on already saved data or models, e.g. when
+    comparing a newly trained model against a baseline in storage.
+
+    Subclasses need to implement the `Artifact.materialize()` API, telling nnbench how to
+    load the desired artifact from a path.
+
+    Parameters
+    ----------
+    path: str | os.PathLike[str]
+        Path to the artifact files.
     """
 
-    pass
+    def __init__(self, path: str | os.PathLike[str]) -> None:
+        # Save the path for later just-in-time materialization.
+        self.path = path
+        self._value: T | None = None
+
+    @classmethod
+    def materialize(cls) -> "Artifact":
+        """Load the artifact from storage."""
+        raise NotImplementedError
+
+    def value(self) -> T:
+        if self._value is None:
+            raise ValueError(
+                f"artifact has not been instantiated yet, "
+                f"perhaps you forgot to call {self.__class__.__name__}.materialize()?"
+            )
+        return self._value
 
 
 @dataclass(frozen=True)
@@ -63,6 +88,19 @@ def __post_init__(self):
         # TODO: Parse interface using `inspect`, attach to the class
 
 
+@dataclass(frozen=True, init=False)
+class Params:
+    """
+    A dataclass designed to hold benchmark parameters. This class is not functional
+    on its own, and needs to be subclassed according to your benchmarking workloads.
+
+    The main advantage over passing parameters as a dictionary is, of course,
+    static analysis and type safety for your benchmarking code.
+    """
+
+    pass
+
+
 def benchmark(
     func: Callable[..., Any] | None = None,
     params: dict[str, Any] | None = None,