fix: reject empty special token strings

[chavalasantosh]

Summary

Reject empty special-token strings when constructing CoreBPE.

An empty special token produces a zero-width match. If that token is allowed
while encoding non-empty text, the match does not advance the encoding cursor,
so the operation does not complete. The configuration currently constructs
without an error.

Changes

• Validate special-token strings in CoreBPE::new_internal.
• Return a clear ValueError through the existing Python constructor wrapper.
• Add a native Rust regression test.
• Add a Python API regression test.

Validation

• cargo test --lib — 3 passed
• pytest tests -q — 34 passed
• cargo fmt --check — passed
• git diff --check — passed

Disclosure

I used AI assistance during investigation and implementation. I reproduced the
behavior, reviewed the complete change, ran the validation above, and take
responsibility for the contribution.