Update on "[llava] Use huggingface LLaVA instead of depending on third-party/LLaVa"

Currently we depend on third-party/LLaVA for llava model definition.
This is hard to use because we have to pull LLaVA in as a git submodule
and install from there. It also breaks a lot of dependency assumptions.

This PR removes `third-party/LLaVA`, in favor of huggingface llava model
definition.

Differential Revision: [D61200610](https://our.internmc.facebook.com/intern/diff/D61200610)

[ghstack-poisoned]

Author: larryliu0820

examples/models/llava/test/test_llava.py

@@ -139,4 +139,4 @@ def test_llava_export(self):
             torch.tensor([new_tokens]), skip_special_tokens=True
         )[0].strip()
         print(outputs)
-        self.assertEqual(new_tokens.size(), 5)
+        self.assertEqual(len(new_tokens), 5)