Validate base64 tokens when loading BPE files

[weivwang]

Summary

Reject malformed Base64 token fields when loading .tiktoken BPE files instead of silently discarding invalid characters.

Python's permissive base64.b64decode turns input such as !!! into b"", so a corrupted vocabulary line could create an empty-byte token and continue loading. Enabling the standard library's strict validation keeps the existing per-line error context while preventing silent vocabulary corruption.

Testing

• .venv/bin/python -m pytest tests/test_load.py -q (2 passed)
• .venv/bin/python -m pytest tests -q --deselect tests/test_encoding.py::test_simple --deselect tests/test_encoding.py::test_basic_encode (31 passed, 4 deselected by existing test selection)
• .venv/bin/python -m compileall -q tiktoken tests/test_load.py
• git diff --check

The two deselected tests fetch tokenizer assets from openaipublic.blob.core.windows.net; they failed before assertions because the local proxy terminated TLS with UNEXPECTED_EOF_WHILE_READING. The new loader tests use local files and cover both valid and invalid Base64 deterministically.