[RAG] Handle TFIDF retriever with pre-trained model

[klshuster]

Patch description
There is currently an issue (#4422) If using a pre-trained RAG/FiD model from the zoo with a TFIDF retriever. This fix includes the query_encoder in the TFIDF retriever state dict.

Testing steps
CI

$ pytest tests/nightly/gpu/test_rag.py  -x
================================================================================ test session starts ================================================================================
platform linux -- Python 3.7.9, pytest-5.3.2, py-1.10.0, pluggy-0.13.1
rootdir: /private/home/kshuster/ParlAI, inifile: pytest.ini
plugins: hydra-core-1.1.0, requests-mock-1.7.0, regressions-2.1.1, datadir-1.3.1
collected 43 items

tests/nightly/gpu/test_rag.py ...........................................                                                                                                     [100%]

============================================================================= slowest 10 test durations =============================================================================
69.95s call     tests/nightly/gpu/test_rag.py::TestRagZooModels::test_bart_rag_dpr_poly
67.92s call     tests/nightly/gpu/test_rag.py::TestRagTfidf::test_rag_token
53.12s call     tests/nightly/gpu/test_rag.py::TestFidZooModels::test_bart_fid_rag_dpr_poly
52.52s call     tests/nightly/gpu/test_rag.py::TestRagDprPoly::test_rag_sequence
50.83s call     tests/nightly/gpu/test_rag.py::TestLoadDPRModel::test_load_dpr
47.06s call     tests/nightly/gpu/test_rag.py::TestRagDprPoly::test_rag_turn
45.98s call     tests/nightly/gpu/test_rag.py::TestRagZooModels::test_bart_rag_sequence
45.39s call     tests/nightly/gpu/test_rag.py::TestRagZooModels::test_bart_rag_turn_do
45.07s call     tests/nightly/gpu/test_rag.py::TestRagZooModels::test_bart_rag_turn_dtt
44.52s call     tests/nightly/gpu/test_rag.py::TestFidZooModels::test_bart_fid_dpr
=================================================================== 43 passed, 24 warnings in 1095.44s (0:18:15) ====================================================================

[jxmsML]

won't the RagAgent inherit this from the TGA?

[klshuster]

i needed to make a change to not use batch.batchsize, since that fails for RAG Sequence.

I suppose the change could actually just be made at the TGA level

[jxmsML]

LGTM! 🚀