test(tokens): cover encode_text without boundaries, decode_text error policies, and modality edge values

[lcampb96]

The existing token tests focus heavily on UTF-8 roundtrip with boundaries=True but leave several paths uncovered. This adds eight tests:

• encode_text with boundaries=False omits BOS/EOS markers
• encode_text byte offset verification (null byte encodes as ID 4)
• decode_text with errors=replace substitutes invalid bytes with U+FFFD
• decode_text with errors=ignore silently drops invalid bytes
• modality(4) correctly classifies the first text token
• modality(259) correctly classifies the last text token before the audio range
• modality(260) correctly classifies the first audio token
• encode_text(, False) yields an empty list

All 63 tests pass locally (pytest tests/test_tokens.py -q).