[Fix] Lazily instantiate backends to avoid unnecessary GPU memory pre-allocations

[kachayev]

Types of changes

• instead of pre-creating backend objects we now only register backend "implementations" (classes)
• instance is created lazily, when requested
• tests are updated accordingly
• OS env flags to disable backends (do not even perform import)
• quickstart mentions how switch backend support off

Motivation and context / Related issue

Closes #516.

How has this been tested (if it applies)

Tests are updated to reflect changes how backends are loaded. Configuration is applied for TF and JAX so they don't allocate GPU memory (which would obviously hurt test performance).

PR checklist

• I have read the CONTRIBUTING document.
• The documentation is up-to-date with the changes I made (check build artifacts).
• All tests passed, and additional code has been covered with new tests.
• I have added the PR and Issue fix to the RELEASES.md file.

[kachayev]

Also, note that there's a somewhat related older PR which should be merged as well: #517

[codecov]

Codecov Report

Merging #520 (a016406) into master (4cf4492) will decrease coverage by 0.10%.
The diff coverage is 90.54%.

Additional details and impacted files

@@            Coverage Diff             @@
##           master     #520      +/-   ##
==========================================
- Coverage   96.08%   95.99%   -0.10%     
==========================================
  Files          65       65              
  Lines       13784    13816      +32     
==========================================
+ Hits        13245    13263      +18     
- Misses        539      553      +14     

[rflamary]

Maybe we should do that in get_backend_list ? I mean this funtiinn should return all bacjen even if the backend was not used in the past... Of course we need to put more detail in the doc of get_backend_list to be clear about the potentiel memory problems

[kachayev]

This depends mainly on the use case get_backend_list was originally dedicated for. get_available_backend_implementations clearly conveys the intention: those are backends "available" to be used. But I'm not sure what would be a use case for get_backend_list except for running tests on all backends at once. If we want get_backend_list to always force instantiation of all them - sure, we can do that. Just being very careful with the documentation.