Passing arbitrary keyword arguments to submodules

[Markus28]

This PR implements a prototype of the proposal #2538. I.e., it allows arbitrary keyword arguments to be passed through SentenceTransformer.encode to the model.

Users may now specify in modules.json what keyword arguments are expected by which module. The SentenceTransformer class no longer inherits from nn.Sequential, but instead from nn.ModuleDict. We implement the forward method ourselves and distribute keyword arguments to the modules.

A model may then provide a modules.json file like this:

[
  {
    "idx": 0,
    "name": "0",
    "path": "",
    "type": "sentence_transformers.models.Transformer",
    "kwargs": ["task", "embedding_dim", "foobar"]
  },
  {
    "idx": 1,
    "name": "1",
    "path": "1_Pooling",
    "type": "sentence_transformers.models.Pooling"
  }
]

and users may use the sentence-transformer model like this:

model = SentenceTransformer('<MODEL>', trust_remote_code=True)
model.encode(['Hello world'], task='sts', embedding_dim=32, foobar=0)

Todo

• These changes will be breaking if there is someone using the .append method, which was previously inherited from nn.Sequential. However, we could also implement it here
• It would make sense to also implement arbitrary kwargs for the tokenizer

[bwanglzu]

@tomaarsen please let us know how do you think about this PR :) thanks

Bo