[Performance]: rollup and time window function. aggregate before sort.

[fengttt]

Is there an existing issue for performance?

• [X] I have checked the existing issues.

Environment

- Version or commit-id (e.g. v0.1.0 or 8b23a93):
- Hardware parameters:
- OS type:
- Others:

Details of Performance

Need to implement efficient processing of rollup and time window function. For both cases, we should run aggregate first (call it phase 1), then sort, and process sorted agg on phase 2. While we actually can hash agg on phase 2 too, in general sort it once is easier/better for these two cases and it is usually what user expect anyway.

Additional information

For example: Have you compared MatrixOne with other databases? If yes, what's their difference?