[TGA] Fix TGA Compatibility with Pytorch>=1.13

[klshuster]

Patch description
Seems pytorch>=1.13 is strict about accessing cpu tensors with gpu items; this PR fixes that.

Addresses #4869

Testing steps
Local testing, + running CI on pytorch 1.13 (which failed before)

$ pytest test_tga.py
=====test session starts =====
platform linux -- Python 3.9.15, pytest-5.3.2, py-1.11.0, pluggy-0.13.1
rootdir: /private/home/kshuster/ParlAI, inifile: pytest.ini
plugins: requests-mock-1.7.0, hydra-core-1.2.0, datadir-1.4.1, regressions-2.1.1
collected 8 items

test_tga.py ........                                                                                                                                                                                                                                                                                                                                                [100%]

===== slowest 10 test durations =====
2.89s call     tests/test_tga.py::TestTGA::test_block_full_context
2.88s call     tests/test_tga.py::TestTGA::test_no_greedy_largebeam
0.59s call     tests/test_tga.py::TestGeneration::test_token_level_loss_logging
0.22s call     tests/test_tga.py::TestGeneration::test_full_context_block
0.21s call     tests/test_tga.py::TestTGA::test_file_inference
0.13s call     tests/test_tga.py::TestGeneration::test_prefix_tokens
0.02s call     tests/test_tga.py::TestGeneration::test_tree_search

(0.00 durations hidden.  Use -vv to show these durations.)
=====8 passed, 13 warnings in 77.97s (0:01:17) =====

[drevicko]

When can we expect to see this in a stable release?

[klshuster]

I'm aiming to get in one more feature before a release compatible with pytorch 1.13 --> so, soon! Keep an eye out for it