More memory optimizations

[JelleAalbers]

This:

• Makes multi-output plugins respect lazy mode (Lazy mailbox #241). Currently divide_outputs doesn't account for lazy mode, so a multi-output plugin could still run away and fetch as much as it likes (potentially blowing up since in lazy mode, the max_messages limits are removed)
• For plugins with multiple inputs, the input controlling the chunk size ('pacemaker') is chosen to be the input that produces the smallest first chunk. Currently we just take the first dependency listed in depends_on.
• OverlapWindowPlugin no longer keeps one chunk back, since it is now OK to yield empty chunks.

These together reduce strax's memory usage in single-core mode by about 15-20%, on top of the similar reduction from #241. Lazy mode now uses about ~30% less memory than non-lazy mode.

The only downside of lazy mode, as noted in #241, is that processing slows down significantly (up to 50% or so) when using single-core lazy mode on a multi-core job/machine. The lazy processing topology assumes you cannot do things in parallel, so when you actually can, it is less efficient.

The multicore processing performance and memory use seem unaffected, although the last two changes listed above do affect multicore processing.