Revert "Load shards as fast as we can"
The current strategy has an unexplained deadlock, which apparently
triggers when read requests happen concurrently with reloads.
Since the GCP load balancer issues a continuing stream of HTTP
requests for /, each of which triggers a read across shards, the
deadlock trigger in production reliably.
This reverts commit e9c19a019ce90ce9ce7021c81c8d0bb10aedeaed.
Change-Id: I5b5234d204a9e9303ab84b098443dd9e9709318f