Add mt.sort support

[qinxuye]

What do these changes do?

This PR adds support for mt.sort which implements the parallel sorting by regular sampling algorithm that mentioned in http://csweb.cs.wfu.edu/bigiron/LittleFE-PSRS/build/html/PSRSalgorithm.html.

I did a little experiment to sort a 100 million 1-d array.

In [1]: n = 100_000_000                                                         

In [2]: import numpy as np                                                      

In [3]: import mars.tensor as mt                                                

In [4]: a = np.random.rand(n)                                                   

In [5]: a.nbytes                                                                
Out[5]: 800000000

In [6]: %%time 
   ...: np.sort(a) 
   ...:  
   ...:                                                                         
CPU times: user 12.1 s, sys: 355 ms, total: 12.4 s
Wall time: 11.4 s

In [9]: at = mt.tensor(a, chunk_size=2_000_000)                                 

In [10]: %%time 
    ...: mt.sort(at).execute(fetch=False) 
    ...:  
    ...:                                                                        
CPU times: user 20.1 s, sys: 1.84 s, total: 21.9 s
Wall time: 3.48 s

It's about 3.3x faster for mt.sort than np.sort on a 100 million length array.

Related issue number

Resolves #828

[hekaisheng]

Great work! some comments left.

[hekaisheng]

LGTM