Checkpoint.docFromText is broken for bsize=None

[jbellis]

It looks like the bsize=None path hasn't been exercised in some time. At the very least it is broken for keep_dims="flatten", because self.doc expects "flatten" to be transformed into "return_mask", and that only happens on the bsize path. It also looks like the rest of the processing in the bsize path is also important for the return value to be transformed correctly but I have not looked more deeply into that.

Here is the exception for the keep_dims issue:

  File "/home/jonathan/Projects/colbert-live/colbert_live/colbert_live.py", line 98, in encode_chunk
    embeddings = self._cp.docFromText([content], keep_dims="flatten", pool_factor=None)
  File "/home/jonathan/miniforge3/envs/colbert-live/lib/python3.10/site-packages/colbert/modeling/checkpoint.py", line 192, in docFromText
    return self.doc(input_ids, attention_mask, keep_dims=keep_dims, to_cpu=to_cpu)
  File "/home/jonathan/miniforge3/envs/colbert-live/lib/python3.10/site-packages/colbert/modeling/checkpoint.py", line 92, in doc
    D = super().doc(*args, **kw_args)
  File "/home/jonathan/miniforge3/envs/colbert-live/lib/python3.10/site-packages/colbert/modeling/colbert.py", line 96, in doc
    assert keep_dims in [True, False, 'return_mask']

[jbellis]

More generally, the options in docFromText don't all play nice together. Eg., pool_factor is ignored unless keep_dims='flatten'.