fix(train): drop unknown tokens during training and inference

[serhiizghama]

Same discrepancy as #353, but on the training path rather than the ONNX one. StaticModel.tokenize drops the unk token, while BaseFinetuneable.tokenize and _prepare_dataset keep it — so a classifier trains and predicts with unk ids still in the sequence, and the to_pipeline() / to_static_model() artifact then drops them. In a small repro built on the mock vocab from the tests, clf.predict_proba and clf.to_pipeline().predict_proba come out 0.13 apart on the same input.

Made both paths mirror StaticModel.tokenize: strip unk first, then truncate. A text that is entirely unknown collapses to an empty sequence, which behaves like an empty string does today and lines up with the zero vector the static model returns for it, so there's no new edge case.

The test checks the trainable tokenizer and the training dataset against StaticModel.tokenize directly; it fails on main for all three tokenizer fixtures.

[codecov]

Codecov Report

❌ Patch coverage is 87.50000% with 1 line in your changes missing coverage. Please review.

Files with missing lines	Patch %	Lines
model2vec/train/base.py	87.50%	1 Missing ⚠️

Files with missing lines	Coverage Δ
model2vec/train/base.py	97.53% <87.50%> (-0.45%)	⬇️

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[stephantul]

Thanks for your PR, nice find! @serhiizghama test coverage is a bit too low, can you make sure it passes? You are missing a test where unk_token_id is None.