[VL] When the regexExpr of the split function is an empty character or a non-ASCII character, fall back to Valina Spark

### Backend

VL (Velox)

### Bug description

**empty character**
```txt
Expression: split('abc','')
Spark: ["a","b","c",""]
Gluten: Reason: (0 vs. 1) split only supports only single-character pattern
```

**non-ASCII character**
```txt
Expression: split('a，b，c','，')
Spark: ["a","b","c"]
Gluten: Reason: (3 vs. 1) split only supports only single-character pattern
```

### Spark version

Spark-3.2.x

### Spark configurations

_No response_

### System information

_No response_

### Relevant logs

```bash
java.lang.RuntimeException: Exception: VeloxRuntimeError
Error Source: RUNTIME
Error Code: INVALID_STATE
Reason: (0 vs. 1) split only supports only single-character pattern
Retriable: False
Expression: patternString.size() == 1
Context: split(n0_0, :VARCHAR)
Top-Level Context: Same as context.
Function: apply
File: ep/build-velox/build/velox_ep/velox/functions/sparksql/SplitFunctions.cpp
```

Start with the Velox split entry point in SplitFunctions.cpp and trace how Gluten dispatches split expressions. Reproduce the empty-pattern and non-ASCII cases against Spark 3.2.x, then verify that unsupported patterns use the Vanilla Spark fallback and match the documented results.

Written by the indexing model from the issue text.