add no_break_image break mode: no partial image at the beginning of a…

[violet-zct]

…n examples

Patch Description
Describe your changes

Testing steps
Describe how you tested your changes

[ArmenAg]

@violet-zct Could we add some tests that cover the behavior we expect? Similar to https://github.com/facebookresearch/metaseq/blob/main/cpu_tests/test_streaming_token_block_dataset.py

[violet-zct]

1. I pushed a cpu test for no_image_break: https://github.com/facebookresearch/metaseq/blob/cm3v2/metaseq/data/document_to_sequence.py#L295
   During my tests, I fix two corner cases in get_document_boundaries() of the cm3 dataset: https://github.com/facebookresearch/metaseq/blob/cm3v2_chunting_dev/metaseq/data/cm3_dataset.py#L208 and #L211, we probably can hit these corner cases when the document is very long.

2. In addition, I found that your get_document_boundaries() can't pass the test: https://github.com/facebookresearch/metaseq/blob/cm3v2_chunting_dev/cpu_tests/test_cm3_dataset.py#L97, unless you remove + 1 in this line: https://github.com/facebookresearch/metaseq/blob/cm3v2_chunting_dev/cpu_tests/test_cm3_dataset.py#L97

3. I fixed image spans by also including the special token at the end of a image.

[ArmenAg]

Tests look great to me. Behavior is as expected. Let's land!