[Python] Bindings for Variant canonical extension type

### Describe the enhancement requested

### **Description:**

The `arrow.parquet.variant` canonical extension type was introduced to C++ in GH-46104 (PR #45375) to represent semi-structured data using a two-buffer (metadata and value) struct layout.

This issue tracks exposing the C++ `VariantExtensionType` class, along with its custom `VariantArray` and `VariantScalar` representations, to PyArrow.

Exposing this type to Python will allow downstream projects to programmatically handle Variant schemas and data.

#### **Proposed Changes:**

1. **Cython Declarations (`python/pyarrow/includes/libarrow.pxd`)**:
* Declare C++ `VariantExtensionType`, `VariantArray`, and `VariantScalar` from `arrow/extension/parquet_variant.h`.
2. **Type & Class Bindings (`python/pyarrow/types.pxi`, `public-api.pxi`)**:
* Expose `VariantType` inheriting from `BaseExtensionType`.
* Register the mapping for `"arrow.parquet.variant"` in `pyarrow_wrap_data_type`.
* Provide a Python-level factory constructor `pyarrow.variant(storage_type)`.
3. **Array & Scalar Bindings (`python/pyarrow/array.pxi`, `python/pyarrow/scalar.pxi`)**:
* Implement `VariantArray` inheriting from `ExtensionArray`.
* Implement `VariantScalar` inheriting from `ExtensionScalar`.
* Provide access to underlying elements (or dictionary representations) via `as_py()` / `to_pydict()`.
4. **Testing**:
* Add comprehensive tests in `python/pyarrow/tests/test_extension_type.py` validating registration, round-trip IPC, pickle compatibility, array building, and scalar inspection.

#### **Dependencies:**

Exposing core type metadata and storage wrapping does not require execution engines, but constructing arrays from Python objects and resolving/decoding values requires the C++ binary encoding/decoding implementations.

Therefore, this issue is blocked by:
* **GH-45946** (PR #50121): [C++][Parquet] Variant decoding
* **GH-45947** (PR #50122): [C++][Parquet] Variant encoding

### Component(s)

Python