fix(embedding): remap unsupported EmbeddingGemma token IDs

[fatkobra]

What does this PR do?

• derive the EmbeddingGemma text-model vocabulary boundary from the tokenizer's base vocabulary
• detect tokenizer IDs that the text-only ONNX embedding table cannot accept
• remap unsupported IDs to the tokenizer's valid <unk> token before invoking ONNX Runtime
• leave valid token batches unchanged
• emit a one-time warning containing the unsupported IDs
• add an offline regression for <image_soft_token>

Root cause

The packaged EmbeddingGemma tokenizer defines <image_soft_token> as added token ID 262144.

The text-only ONNX export's embedding matrix has 262,144 rows, so valid indices end at 262143.

EmbeddinggemmaONNX.__call__() previously passed the tokenizer output directly to session.run(). When source text contained the literal <image_soft_token>, ONNX Runtime attempted to gather row 262144 and aborted the complete embedding batch.

Implementation

The new helper obtains the text-model boundary with:

tokenizer.get_vocab_size(with_added_tokens=False)

How to test

python3 -m ruff format mempalace/embedding.py tests/test_embeddinggemma.py
python3 -m ruff format --check .
python3 -m ruff check .
python3 -m pytest tests/test_embeddinggemma.py \
  -k "out_of_range_added_token_is_remapped_to_unknown" \
  -q \
  -rs
python3 -m pytest tests/test_embeddinggemma.py -q -rs
python3 -m pytest \
  tests/test_embeddinggemma.py \
  tests/test_embedding.py \
  tests/test_embedding_wrapper.py \
  -q \
  -rs
python -m pytest tests/ -v

Checklist

• Tests pass (python -m pytest tests/ -v)
• No hardcoded paths
• Linter passes (ruff check .)