[Fix] Preserve native BPE training with protected separators

[baihchou8787]

Description

Fixes GraphBPE so protected component separators split BPE input sequences without forcing backend=cpp or backend=auto onto the Python training path. Explicit cpp now fails during fit when the native extension is unavailable.

The native bridge accepts an initial vocabulary floor so merge IDs cannot collide with protected token IDs excluded from training segments.

Checklist

• The PR title starts with [Fix]
• Changes are complete
• All changes have test coverage
• Code is well-documented
• Examples are not affected
• Related issue is referred in this PR

Changes

• Segment protected tokens before BPE training and encoding.
• Keep native training available for protected GraphTokenizer separators.
• Add regressions for native dispatch and fail-fast behavior.

Verification

TL_BACKEND=torch /local/wrq/gammagl-graphtokenizer/bin/python -m pytest tests/transforms/test_graph_tokenizer.py tests/transforms/test_graph_bpe_cpp_install.py -q

Result: 36 passed.