fix: correctly decode numpy UCS-4 arrays instead of treating as UTF-8

[VirajMishra1]

PyArrayUnicode in the Python bindings passes raw numpy dtype='U' array bytes directly to std::str::from_utf8(). But numpy stores Unicode strings as UCS-4 (UTF-32LE on little-endian), not UTF-8. This causes:

• ASCII strings: silent data corruption -- embedded null bytes between every character (e.g. "hello" becomes "h\0\0\0e\0\0\0l\0\0\0l\0\0\0o\0\0\0")
• Non-ASCII strings: ValueError('invalid utf-8 sequence') for any input containing accented characters, CJK, emoji, etc.

The commented-out code directly below the broken path (using PyUnicode_FromKindAndData with PyUnicode_4BYTE_KIND) was the correct approach but was replaced with the from_utf8 call.

Fix

Decode 4-byte chunks as UTF-32LE code points (u32::from_le_bytes), filter null padding, convert via char::from_u32, and collect to String. This correctly handles numpy's UCS-4 encoding for all Unicode inputs.

Test

Added test_numpy_unicode_ucs4 in bindings/python/tests/bindings/test_tokenizer.py covering ASCII, CJK, and emoji inputs through numpy dtype='U' arrays.