[Feature request] `finalize_reducer` support vectorization and more op type

Now `finalize_reducer` only support op in ("sum", "max", "min"). (See `tilelang/language/allocate.py` line 146)

We have `reduce_abssum` and `reduce_absmax`, which can use in `finalize_reducer`.

What's more, `finalize_reducer` now don't support vectorization. This can lead to inefficient kernel performance in certain cases.

Start at tilelang/language/allocate.py around line 146 and trace how finalize_reducer handles the current sum, max, and min operations. Compare that path with reduce_abssum and reduce_absmax, then identify how vectorization is represented in the surrounding reducer logic. Done means the additional operations and vectorized finalize_reducer use cases are supported without the current performance limitation.

Written by the indexing model from the issue text.