[EPIC] Improve NUMA-awareness

### Is your feature request related to a problem or challenge?

This is a list to check where DataFusion w.r.t. NUMA-awareness and where we can improve.

I will list to some other tickets.

- https://github.com/apache/datafusion/issues/21719
- [ ] Support morsel-splitting in Parquet, other datasources for better parallelisation / distribution / prefetching
- [x] Read parquet metadata to use `ParquetMetaDataPushDecoder`
- [ ] Support/use pinning threads to cores to minimize context switches / cache misses)
- [ ] Support reading into thread-local storage using async io / `io_uring` instead of using `spawn_blocking`
- [ ] Split/decode row groups into sizes that fit well in cache (e.g. ~RecordBatch)
- [ ] Better query execution pipelining / scheduling to minimize context switching / optimize data locality

### Describe the solution you'd like

_No response_

### Describe alternatives you've considered

_No response_

### Additional context

_No response_

Start with the linked issue 21719 and review the NUMA-awareness checklist in this issue. The listed areas include Parquet morsel-splitting, ParquetMetaDataPushDecoder, thread pinning, async I/O, cache-sized row groups, and query scheduling; the work is complete only when a specific improvement is selected, implemented, and its checklist item can be marked done.

Written by the indexing model from the issue text.