Add a benchmark for dask.array.block comparing it to numpy and straight copy

[hmaarrfk]

@mrocklin I guess it wasn't as slow as I remember :S.

Dask is doing better than numpy, at least for now 👯‍♀️

===== ========== ================= =============== ========================= ============= ========== ==========
--                                                       mode                                                   
----- ----------------------------------------------------------------------------------------------------------
  n     block     block optimized   block persist   block optimized persist   concatenate   np_block   np_copy  
===== ========== ================= =============== ========================= ============= ========== ==========
  1    4.50±0ms       2.59±0ms         1.82±0ms             2.46±0ms            3.20±0ms    139±0μs    62.5±0μs 
  10   3.01±0ms       2.13±0ms         2.69±0ms             2.73±0ms            4.41±0ms    1.41±0ms   427±0μs  
 100   164±0ms        163±0ms          1.84±0ms             1.72±0ms            139±0ms     440±0ms    141±0ms  
===== ========== ================= =============== ========================= ============= ========== ==========

[jakirkham]

Is this ready to merge?

[hmaarrfk]

@jakirkham is there a better way to optimize the graph here?

[jakirkham]

This is a pretty typical strategy for optimizing. Does a few different things for Dask Arrays in particular.

[hmaarrfk]

Thanks. I would like to get a

1. Typical dask user (i.e. no optimization called)
2. Call to block with recommended optimizations to the graph.
3. Calling persist with the above combos.

[jakirkham]

Calling persist will trigger the standard Dask Array graph optimizations in either case.

[hmaarrfk]

Got it, I'll remove that "mode".