Find good defaults for reduction parameters

[tomwhite]

There are a couple of parameters we can set:

• split_every, which controls the number of blocks in each stage of tree reduce (defaults to 4, but set to 10 in the quadratic means benchmark)
• max_total_num_input_blocks, which controls how many blocks can be read by one task when doing optimization (not enabled by default, but set to 40 in the quadratic means benchmark)

This comment discusses the trade off to optimize: #331 (comment)