[SPARK-3478] [PySpark] Profile the Python tasks

[davies]

This patch add profiling support for PySpark, it will show the profiling results
before the driver exits, here is one example:

============================================================
Profile of RDD<id=3>
============================================================
         5146507 function calls (5146487 primitive calls) in 71.094 seconds

   Ordered by: internal time, cumulative time

   ncalls  tottime  percall  cumtime  percall filename:lineno(function)
  5144576   68.331    0.000   68.331    0.000 statcounter.py:44(merge)
       20    2.735    0.137   71.071    3.554 statcounter.py:33(__init__)
       20    0.017    0.001    0.017    0.001 {cPickle.dumps}
     1024    0.003    0.000    0.003    0.000 t.py:16(<lambda>)
       20    0.001    0.000    0.001    0.000 {reduce}
       21    0.001    0.000    0.001    0.000 {cPickle.loads}
       20    0.001    0.000    0.001    0.000 copy_reg.py:95(_slotnames)
       41    0.001    0.000    0.001    0.000 serializers.py:461(read_int)
       40    0.001    0.000    0.002    0.000 serializers.py:179(_batched)
       62    0.000    0.000    0.000    0.000 {method 'read' of 'file' objects}
       20    0.000    0.000   71.072    3.554 rdd.py:863(<lambda>)
       20    0.000    0.000    0.001    0.000 serializers.py:198(load_stream)
    40/20    0.000    0.000   71.072    3.554 rdd.py:2093(pipeline_func)
       41    0.000    0.000    0.002    0.000 serializers.py:130(load_stream)
       40    0.000    0.000   71.072    1.777 rdd.py:304(func)
       20    0.000    0.000   71.094    3.555 worker.py:82(process)

Also, use can show profile result manually by sc.show_profiles() or dump it into disk
by sc.dump_profiles(path), such as

>>> sc._conf.set("spark.python.profile", "true")
>>> rdd = sc.parallelize(range(100)).map(str)
>>> rdd.count()
100
>>> sc.show_profiles()
============================================================
Profile of RDD<id=1>
============================================================
         284 function calls (276 primitive calls) in 0.001 seconds

   Ordered by: internal time, cumulative time

   ncalls  tottime  percall  cumtime  percall filename:lineno(function)
        4    0.000    0.000    0.000    0.000 serializers.py:198(load_stream)
        4    0.000    0.000    0.000    0.000 {reduce}
     12/4    0.000    0.000    0.001    0.000 rdd.py:2092(pipeline_func)
        4    0.000    0.000    0.000    0.000 {cPickle.loads}
        4    0.000    0.000    0.000    0.000 {cPickle.dumps}
      104    0.000    0.000    0.000    0.000 rdd.py:852(<genexpr>)
        8    0.000    0.000    0.000    0.000 serializers.py:461(read_int)
       12    0.000    0.000    0.000    0.000 rdd.py:303(func)

The profiling is disabled by default, can be enabled by "spark.python.profile=true".

Also, users can dump the results into disks automatically for future analysis, by "spark.python.profile.dump=path_to_dump"

This is bugfix of #2351 cc @JoshRosen

[AmplabJenkins]

Test FAILed.
Refer to this link for build results (access rights to CI server needed):
https://amplab.cs.berkeley.edu/jenkins//job/SparkPullRequestBuilder/20903/

[SparkQA]

QA tests have started for PR 2556 at commit e68df5a.

• This patch merges cleanly.

[SparkQA]

QA tests have finished for PR 2556 at commit e68df5a.

• This patch passes unit tests.
• This patch merges cleanly.
• This patch adds no public classes.

[JoshRosen]

I've merged this. Thanks for the fix!