[FIX][AUTOTVM] Make autotvm work with spawn

[tkonolige]

Like #6671 this PR fixes autotvm when using the spawn start method for multiprocessing. I've added some tests to make sure that things work with spawn in the CI. I've also added notices to the tutorials letting users know that they will not run on windows or macOS.

Here are performance results for tutorials/autotvm/tune_conv2d_cuda.py with n_trials = 100.

This PR

XGB iter:   0	tr-a-recall@64: 0.661270	tr-map: 0.100000
XGB iter:  25	tr-a-recall@64: 0.754100	tr-map: 1.000000
XGB stopped. Best iteration: [29] tr-a-recall@64:0.75971	tr-map:1.00000
XGB train: 1.08	obs: 64	error: 43	n_cache: 64
SA iter: 50	last_update: 49	max-0: 5.39	max-1: 6.08	temp: 0.90	elapsed: 3.41
SA iter: 100	last_update: 98	max-0: 5.72	max-1: 6.25	temp: 0.80	elapsed: 6.82
SA iter: 150	last_update: 148	max-0: 5.89	max-1: 6.51	temp: 0.70	elapsed: 10.38
SA iter: 200	last_update: 197	max-0: 5.97	max-1: 6.51	temp: 0.60	elapsed: 14.18
SA iter: 250	last_update: 248	max-0: 6.09	max-1: 6.51	temp: 0.50	elapsed: 18.02
SA iter: 300	last_update: 298	max-0: 6.15	max-1: 6.51	temp: 0.40	elapsed: 21.72
SA iter: 350	last_update: 348	max-0: 6.17	max-1: 6.51	temp: 0.30	elapsed: 25.30
SA iter: 400	last_update: 398	max-0: 6.21	max-1: 6.51	temp: 0.20	elapsed: 28.79
SA iter: 450	last_update: 447	max-0: 6.24	max-1: 6.51	temp: 0.10	elapsed: 32.20
SA iter: 500	last_update: 493	max-0: 6.25	max-1: 6.51	temp: 0.00	elapsed: 35.35
SA iter: 500	last_update: 493	elapsed: 35.35


main

XGB iter:   0	tr-a-recall@64: 0.610293	tr-map: 0.500000
XGB iter:  25	tr-a-recall@64: 0.671535	tr-map: 1.000000
XGB stopped. Best iteration: [29] tr-a-recall@64:0.67466	tr-map:1.00000
XGB train: 1.02	obs: 64	error: 50	n_cache: 64
SA iter: 50	last_update: 49	max-0: 5.41	max-1: 6.23	temp: 0.90	elapsed: 3.41
SA iter: 100	last_update: 98	max-0: 5.58	max-1: 6.23	temp: 0.80	elapsed: 6.94
SA iter: 150	last_update: 145	max-0: 5.72	max-1: 6.37	temp: 0.70	elapsed: 10.61
SA iter: 200	last_update: 198	max-0: 5.76	max-1: 6.37	temp: 0.60	elapsed: 14.25
SA iter: 250	last_update: 247	max-0: 5.83	max-1: 6.45	temp: 0.50	elapsed: 17.87
SA iter: 300	last_update: 298	max-0: 6.03	max-1: 6.45	temp: 0.40	elapsed: 21.81
SA iter: 350	last_update: 349	max-0: 6.07	max-1: 6.45	temp: 0.30	elapsed: 25.57
SA iter: 400	last_update: 397	max-0: 6.11	max-1: 6.45	temp: 0.20	elapsed: 28.99
SA iter: 450	last_update: 444	max-0: 6.13	max-1: 6.45	temp: 0.10	elapsed: 32.60
SA iter: 500	last_update: 489	max-0: 6.14	max-1: 6.45	temp: 0.00	elapsed: 35.76
SA iter: 500	last_update: 489	elapsed: 35.76

This is on a 64-core x86 ubuntu system with python 3.6.9. Performance is the same.

[merrymercy]

@tkonolige Thanks! It is merged.

[merrymercy]

@tkonolige This PR introduces cloudpickle as a dependency.

Could you add this to https://github.com/apache/tvm/blob/main/docs/install/from_source.rst#python-dependencies and all autotvm tutorials (

tvm/tutorials/autotvm/tune_relay_cuda.py

Line 50 in 878a0a9

# pip3 install --user psutil xgboost tornado

and other places)?

Or can we do not introduce cloudpicke when fork is used?