Fix new boost libs failure in cache-lib mode and add test to cover collect_dependent_libs.sh

[guoyuhong]

What do these changes do?

This PR #3574 breaks the script that builds ray using cached libs.
This PR does the following things:

1. Fix the building failure introduced by Fix TensorFlow and PyTorch compatibility #3574 .
   • The cached boost lib does not realize there are version mismatch, so this PR add a version check. If there is version mismatch, correct boost will be downloaded and built.
   • The cached boost lib does not contain the new Boost::Thread lib.
2. Add a test to cover collect_dependent_libs.sh and cover the build-with-cache-lib mode.

Related issue number

#3521

[AmplabJenkins]

Test FAILed.
Refer to this link for build results (access rights to CI server needed):
https://amplab.cs.berkeley.edu/jenkins//job/Ray-PRB/10360/
Test FAILed.

[guoyuhong]

The valgrind test time grows from about 18 minutes to around 24 minutes.

[guoyuhong]

@pcmoritz Please take a look when you are spared.

[AmplabJenkins]

Test PASSed.
Refer to this link for build results (access rights to CI server needed):
https://amplab.cs.berkeley.edu/jenkins//job/Ray-PRB/10414/
Test PASSed.

[guoyuhong]

@robertnishihara Do you have any comments on this PR?

[AmplabJenkins]

Test FAILed.
Refer to this link for build results (access rights to CI server needed):
https://amplab.cs.berkeley.edu/jenkins//job/Ray-PRB/10543/
Test FAILed.