DEV: Add ability to run memory benchmarks on macOS

[alugowski]

Some benchmarks such as io_matlab can only run on Linux because:

• setrlimit may throw on macOS
• memory total/available stats are read from /proc
• a subprocess' peak memory usage is a max of polled RSS size in /proc

This PR addresses all three such that benchmarks that measure memory usage also work on macOS.

• handle a common exception setrlimit raises on macOS
• memory total/available stats are fetched from psutil
• subprocess' peak memory usage is read from the rusage struct returned by os.wait4()

The key contribution is that final bit. os.wait4() waits on a process and returns its rusage struct. This struct contains the peak RSS usage by the child process.

os.wait4() is available on Unix systems, including Linux and macOS. So this method adds macOS support. The other upside is that this is a true max, not a max of polling samples.

[rgommers]

This looks great, thanks a lot @alugowski!

I tested on macOS, and indeed it fixes the loadmat/savemat memory benchmarks (results report peak_mem / size, so the smaller numbers for larger size make sense):

[ 50.00%] ··· ====== ==================== ====================
              --                     compressed               
              ------ -----------------------------------------
               size          True                False        
              ====== ==================== ====================
                1M        51.658752             51.11808      
               10M        5.9834368            5.9342848      
               100M       1.50798336           1.5015936      
               300M   1.1664861866666667   1.1606971733333333 
              ====== ==================== ====================

[100.00%] ··· io_matlab.MemUsage.track_savemat                                                                         ok
[100.00%] ··· ====== =================== ====================
              --                    compressed               
              ------ ----------------------------------------
               size          True               False        
              ====== =================== ====================
                1M        51.953664           50.085888      
               10M        9.4420992           6.8354048      
               100M       4.42466304          2.49004032     
               300M   4.048431786666667   2.1637802666666666 
              ====== =================== ====================

All LGTM, so in it goes.