Ability to set max_seq_len to the SentenceTransformers components

[bglearning]

Is your feature request related to a problem? Please describe.
The SentenceTransformer models have a max_seq_len attribute.

In theory, we could set it with the model_max_length in tokenizer_kwargs which then eventually should set that attribute here.

However, it seems to be unreliable. We saw it with bge-m3 but could also be the case for other models. (Colab)
This can result in OOM error when embedding.

Update: found the "issue". The max_seq_length is read into the kwargs from this config json at this point in _load_sbert_model and thus the max_seq_length is not None here and so it doesn't use model_max_length from the tokenizer_kwargs.

Describe the solution you'd like
Would be good to have the ability to set the max_seq_len in the (currently three) SentenceTransformers components as in v1.

We could possibly also intercept the tokenizer_kwargs and use model_max_length from it if it's set.