PyTorch: Release Python GIL when yielding in foreground thread

[maxhgerlach]

Checklist before submitting

• Did you read the contributor guide?
• Did you update the docs?
• Did you write any tests to validate this change?
• Did you update the CHANGELOG, if this change affects users?

Description

Fixes #3352. I've run the test suite from tests/parallel/test_torch.py 15 times in the environment described in the issue and did not encounter any hang.

I hope releasing the GIL here is not unsafe in any way.

[github-actions]

Unit Test Results

     802 files   -   28       802 suites   - 28   9h 8m 17s ⏱️ - 9m 32s
     717 tests ±    0       672 ✔️ ±    0       45 💤 ±    0  0 ❌ ±0 
17 324 runs   - 664  12 238 ✔️  - 406  5 086 💤  - 258  0 ❌ ±0 

Results for commit dcf942c. ± Comparison against base commit 976a879.

[github-actions]

Unit Test Results (with flaky tests)

     902 files   -      54       902 suites   - 54   10h 34m 46s ⏱️ + 17m 57s
     717 tests ±       0       671 ✔️ +    2       45 💤 ±    0  1 ❌  - 2 
19 772 runs   - 1 060  13 846 ✔️  - 606  5 922 💤  - 454  4 ❌ ±0 

For more details on these failures, see this check.

Results for commit dcf942c. ± Comparison against base commit 976a879.

[maxhgerlach]

One of the GPU head test suites on Buildkite failed because of some Docker trouble; it does not look related.