PlanLoader addition into planner (meta-pytorch#3355)

Summary:
Pull Request resolved: meta-pytorch#3355

**Summary:**

*   Added PlanLoader abstract base class to enable loading pre-computed sharding plans from stored locations within planner.
*   Supports two key scenarios:
    1.  Reusing previously computed and stored sharding plans to avoid regeneration costs
    2.  Using sharding plans from previous runs as starting points for iterative improvements
*   Defines two abstract methods:
    *   `load()`: Returns a dictionary mapping sharding option hashes to ShardingOption objects
    *   `plan_validation_str()`: Provides validation string for plan integrity checks
*   Part of the broader effort to improve planner UX and reliability by enabling plan persistence and reuse across training runs

Differential Revision: D81571293

Author: aliafzal

torchrec/distributed/planner/types.py

@@ -995,6 +995,40 @@ def log(
         ...
 
 
+class PlanLoader(abc.ABC):
+    """
+    Retrieves a pre-computed sharding plan from its stored location. This is useful in two scenarios:
+        1. To utilize a specific sharding plan that was previously computed and stored, saving the cost of re-generating the plan
+        2. To use a sharding plan from previous runs as a starting point for the next run, allowing for improvement over time.
+    """
+
+    @abc.abstractmethod
+    def load(
+        self,
+    ) -> Optional[Dict[int, ShardingOption]]:
+        """
+        Load sharding plan from its stored location.
+
+        Returns:
+            Dict[int, ShardingOption]: loaded sharding plan. key is hash of sharding option to map to sharding option with enumerated sharding option.
+        """
+        ...
+
+    @abc.abstractmethod
+    def plan_context_hash(
+        self,
+    ) -> Optional[str]:
+        """
+        Input context hash of a sharding plan.
+
+        Returns:
+            str: hash of sharding plan context.
+        """
+        ...
+
+    ...
+
+
 @dataclass
 class CriticalPathEstimate:
     comms_estimate: float