Add observation_function for full custom control over the observation

Lets a game define how its state becomes an observation as a plain
Python function (module:attr), used identically by GameEnvironment
(training) and TrainedPolicy (inference) instead of two independently
maintained encoding paths. Removes the egocentric/egocentric_player/
egocentric_radius flags — cropping is now something an
observation_function does itself by calling egocentric_board(), and
extras_size is discovered from one sampled observation instead of
being configured via observe_state_sizes.

retro_gamer/cli.py

@@ -13,6 +13,12 @@ from retro_gamer.trainer import DQNTrainer, DEFAULTS, MODEL_KEYS
 @click.group()
 def cli():
     """Train and run RL agents for retro games."""
+    # Running the installed console script puts its own directory on sys.path,
+    # not the caller's cwd — but observation_function/game modules are
+    # typically plain files in the directory you ran retro-gamer from.
+    cwd = str(Path.cwd())
+    if cwd not in sys.path:
+        sys.path.insert(0, cwd)
 
 
 # ---------------------------------------------------------------------------
@@ -79,8 +85,9 @@ def create(game, output, **hyperparams):
         raise click.ClickException(str(e))
 
     game_factory = _load_factory(game_config)
-    g = game_factory()
-    metadata.board_size = g.board_size
+    if metadata.board_size is None:
+        g = game_factory()
+        metadata.board_size = g.board_size
 
     metadata.validate()
 
@@ -108,6 +115,8 @@ def create(game, output, **hyperparams):
     else:
         click.echo(f"  characters  : (will be auto-discovered during training)")
     click.echo(f"  architecture: {'CNN (spatial)' if metadata.spatial else 'MLP (non-spatial)'}")
+    if metadata.observation_function:
+        click.echo(f"  observation : {metadata.observation_function} (custom)")
 
 
 # ---------------------------------------------------------------------------