Use ParquetPushDecoder in ParquetOpener (#20839)

## Which issue does this PR close?

- Closes #20841

## Rationale for this change

We want to split IO and CPU to allow for more (NUMA-aware) parallelism
and utilizing IO and CPU better.
This allows for e.g. more coalescing, prefetching, parallel IO, more
parallel / incremental decoding etc.
Also this allows doing morsels only on a CPU level and not doing IO
multiple times for each morsel.

## What changes are included in this PR?

Just refactor `ParquetOpener` to use `ParquetPushDecoder`. I used claude
to rewrite it and to keep changes small.

## Are these changes tested?

Existing tests. Nothing should change, the arrow-rs code also uses
`ParquetPushDecoder`.

## Are there any user-facing changes?

---------

Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>

Author: Dandandan