[parquet] reduce the time spent in CachedArrayReader

**Is your feature request related to a problem or challenge? Please describe what you are trying to do.**

I am profiling clickbench query 26 with predicate pushdown enabled as part of
- https://github.com/apache/datafusion/issues/3463

```shell
samply record -- /Users/andrewlamb/Software/datafusion2/target/profiling/datafusion-cli -f q.sql > /dev/null 2>&1
```

```sql
SELECT "SearchPhrase" FROM hits WHERE "SearchPhrase" <> '' ORDER BY "EventTime", "SearchPhrase" LIMIT 10;
```

While looking at the profile, I noticed that 3% of the time is spent concatenating in the cached array reader

I believe the call is here:
https://github.com/apache/arrow-rs/blob/814ee4227c01fce478bdd3594dd156250286b46e/parquet/src/arrow/array_reader/cached_array_reader.rs#L333

**Describe the solution you'd like**
I would like to make this faster

**Describe alternatives you've considered**
I think we can use the [`BatchCoalescer`](https://docs.rs/arrow/latest/arrow/compute/struct.BatchCoalescer.html) for this task and potentially save at least one copy

**Additional context**

Start with parquet/src/arrow/array_reader/cached_array_reader.rs around line 333 and reproduce the ClickBench query using the provided samply command. Investigate whether BatchCoalescer can reduce concatenation cost or copying, then compare profiling results while preserving the query output.

Written by the indexing model from the issue text.