This adds a "flush_interval" top-level config to the table config,
telling how much time should pass between flushes.

This allows multiple writes to be combined into a single flush,
reducing disk bandwidth, disk lifetime, and for many workloads, write
throughput.

Eviction now works such that if a table shard is using too much memory
(and none are evictable bufs), the whole shard initiates a flush.
(Previously, it _waited_ for enough active flushes to complete -- now
it must initiate the flush.)  It can't yet incrementally evict bufs,
because that is trickier to implement correctly than you'd think.