refactor: simplify training and inference interfaces

- avoid constructing model_fn more than once when reading config
- keep inference package exports focused on public entry points
- rename extra strategy arguments to strategy_kwargs

tests/inference/test_cache.py

@@ -2,7 +2,7 @@
 
 import torch
 
-from astrai.inference import (
+from astrai.inference.cache import (
     Allocator,
     KVStorage,
     PagePool,