arrow

GH-41334: [C++][Acero] Use per-node basis temp vector stack to mitigate overflow (#41335)



### Rationale for this change

The risk of temp vector stack overflow still exists as described in #41334 . Many people have agreed on a per-node basis approach:
> 1) it doesn't introduce more performance penalty than shared stack; 2) it can mitigate the overflow in a natural way, i.e., expanding the stack size linear to the number of nodes; 3) it requires no more complexity to the existing stack implementation.

The full (but long) story is also revealed in the subsequent discussion of this PR. Feel free to scroll down.

### What changes are included in this PR?

1. Change the current shared (per-thread) temp vector stack usage to per-node basis.
2. Make the stack size required by each stack user more explicit.

### Are these changes tested?

UT included.

### Are there any user-facing changes?

None.

* GitHub Issue: #41334

Authored-by: Ruoxi Sun <zanmato1984@gmail.com>
Signed-off-by: Antoine Pitrou <antoine@python.org>