[mlir][Vector] Switch ExtractOp to the declarative assembly format

This is a bit awkward since ExtractOp allows both `f32` and
`vector<1xf32>` results for a scalar extraction. Allow both, but make
inference return the scalar to make this as NFC as possible.