Add dialogue data collator tests (#2485)

Add dialogue data collator unit test. Things to note on this PR:
- is it correct that we mask the last occurance of `<|endoftext|>` of
the assistant? See the example in the test, there will be one occurance
where we have this token and one where there is none. See the todo in
the code.
- I built a dummy tokenizer from the pythia-70m one using `tokenizer =
old_tokenizer.train_new_from_iterator(training_iter, vocab_size)` to
keep the size minimal. Just trained on the text that appears in the
test.

model/model_training/tests/resources/data_collator/special_tokens_map.json

@@ -0,0 +1,14 @@
+{
+  "additional_special_tokens": [
+    "<|prompter|>",
+    "<|assistant|>",
+    "<|system|>",
+    "<|prefix_begin|>",
+    "<|prefix_end|>"
+  ],
+  "bos_token": "<|endoftext|>",
+  "eos_token": "<|endoftext|>",
+  "pad_token": "<|padding|>",
+  "sep_token": "<|endoftext|>",
+  "unk_token": "<|endoftext|>"
+}