Fix and test llamacpp

[pseudotensor]

Now properly supports:

• GPT4All gptj: downloads to cache, no local file choice
• GPT4All llama (version 2 quant), downloads to cache, no local file choice
• llama-cpp-python's llama (version 3 quant)

Saw odd garbling from model had downloaded myself from model explorer, so sticking to internal cache downloader for GPT4All.

gptj from GPT4all still has the "no output" issue, even natively. So best to stick to llama.cpp version IMO, e.g. by downloading a version 3 quantization llama.cpp model, or build one yourself as in FAQ.md, and then doing:

python generate.py --base_model='llama' --langchain_mode=UserData --user_path=user_path --score_model=None

Tests:

============================================================================================================ short test summary info ============================================================================================================
FAILED tests/test_manual_test.py::test_chat_context - NotImplementedError: MANUAL TEST FOR NOW
FAILED tests/test_manual_test.py::test_upload_one_file - NotImplementedError: MANUAL TEST FOR NOW -- do and ask query of file
FAILED tests/test_manual_test.py::test_upload_multiple_file - NotImplementedError: MANUAL TEST FOR NOW -- do and ask query of files
FAILED tests/test_manual_test.py::test_upload_url - NotImplementedError: MANUAL TEST FOR NOW -- put in URL box https://github.com/h2oai/h2ogpt/ (and ask what is h2ogpt?). Ensure can go to source links
FAILED tests/test_manual_test.py::test_upload_arxiv - NotImplementedError: MANUAL TEST FOR NOW -- paste in arxiv:1706.03762 and ask who wrote attention paper. Ensure can go to source links
FAILED tests/test_manual_test.py::test_upload_pasted_text - NotImplementedError: MANUAL TEST FOR NOW -- do and see test code for what to try
FAILED tests/test_manual_test.py::test_no_db_dirs - NotImplementedError: MANUAL TEST FOR NOW -- Remove db_dirs, ensure can still start up and use in MyData Mode.
FAILED tests/test_manual_test.py::test_upload_unsupported_file - NotImplementedError: MANUAL TEST FOR NOW -- e.g. json, ensure error correct and reasonable, no cascades
FAILED tests/test_manual_test.py::test_upload_to_UserData_and_MyData - NotImplementedError: MANUAL TEST FOR NOW Upload to each when enabled, ensure no failures
FAILED tests/test_manual_test.py::test_chat_control - NotImplementedError: MANUAL TEST FOR NOW save chat, select chats, clear chat, export, import, etc.
FAILED tests/test_manual_test.py::test_subset_only - NotImplementedError: MANUAL TEST FOR NOW UserData, Select Only for subset, then put in whisper.  Ensure get back only chunks of data with url links to data sources.
=========================================================================== 11 failed, 60 passed, 25 skipped, 1 xfailed, 1 xpassed, 2 warnings in 1198.21s (0:19:58) ============================================================================
(h2ollm) jon@pseudotensor:~/h2o-llm$ 

[pseudotensor]

This fixes many problems, but still llama/gpt4all do poorly with context from langchain. Will improve next.