fix: disable automatic eviction of diskcache at aggregation

Python diskcache, which we use for aggregating trainer updates, have its
own automatic eviction policy, depending on its size_limit and
cull_limit value. Updated it to disable automatic eviction, and added a
logger debug line that tells you the number of updates in the cache.

lib/python/flame/mode/horizontal/syncfl/top_aggregator.py

@@ -93,7 +93,11 @@ def internal_init(self) -> None:
         self.metrics = dict()
 
         # disk cache is used for saving memory in case model is large
+        # automatic eviction of disk cache is disabled with cull_limit 0
         self.cache = Cache()
+        self.cache.reset('size_limit', 1e15)
+        self.cache.reset('cull_limit', 0)
+
         self.optimizer = optimizer_provider.get(
             self.config.optimizer.sort, **self.config.optimizer.kwargs
         )
@@ -157,6 +161,8 @@ def _aggregate_weights(self, tag: str) -> None:
                 # save training result from trainer in a disk cache
                 self.cache[end] = tres
 
+        logger.debug(f"received {len(self.cache)} trainer updates in cache")
+
         # optimizer conducts optimization (in this case, aggregation)
         global_weights = self.optimizer.do(
             deepcopy(self.weights),