Apply TrainingArguments.max_length during embedding training

[robbiebusinessacc]

Fixes #561.

TrainingArguments.max_length is documented as "the maximum token length a
tokenizer can generate," and it is applied when fitting the classifier head
(Trainer.train_classifier passes it to SetFitModel.fit, which builds the
dataloader with that length). It was never applied to the embedding phase,
though: Trainer.train_embeddings finetunes the SentenceTransformer body via
the underlying trainer without ever setting the body's truncation length, so
max_length had no effect there — matching the report that batches were padded
to the longest example instead of being truncated.

This sets the body's max_seq_length to max_length for the embedding phase
(clamped to the model's maximum, mirroring the existing clamp in
_prepare_dataloader) and restores the original value afterwards, so
inference-time encoding is unaffected — keeping the same non-persistent
behavior the classifier phase already has.

Test: test_trainer_max_length_applied_to_embedding_phase asserts the body is
truncated to max_length during the embedding phase and restored afterwards.
It fails on main (the body stays at its default 100) and passes with this
change.