fix(py): make encode_with_unstable handle surrogates like encode (#541)

[MasterOfLogic1]

Summary

Fixes #541.

Encoding.encode and Encoding.encode_ordinary already catch UnicodeEncodeError from the Rust BPE layer and retry after a UTF-16 surrogatepass + replace round-trip. Encoding.encode_with_unstable did not, so the same inputs that worked through encode raised UnicodeEncodeError.

This PR mirrors the same try/except/repair pattern in encode_with_unstable so all three encode paths accept the same surrogate inputs.

Repro (before)

import tiktoken

enc = tiktoken.get_encoding("cl100k_base")

enc.encode("\ud83d\udc4d")               # works 
enc.encode_with_unstable("\ud83d\udc4d") # UnicodeEncodeError before but works 

enc.encode("\ud83d")                     # works
enc.encode_with_unstable("\ud83d")       # UnicodeEncodeError

Changes

Wrap self._core_bpe.encode_with_unstable(...) in the same surrogate repair logic used by encode() in tiktoken/core.py
Add test_encode_with_unstable_surrogate_pairs in tests/test_encoding.py

Test Plan

def test_encode_with_unstable_surrogate_pairs():
    enc = tiktoken.get_encoding("cl100k_base")

    # would raise UnicodeEncodeError before the fix in core.py
    enc.encode_with_unstable("\ud83d\udc4d")
    enc.encode_with_unstable("\ud83d")

    assert enc.encode("\ud83d\udc4d") == enc.encode("👍")
    assert enc.encode("\ud83d") == enc.encode("�")