GH-55281: [C++][Gandiva] Release arena memory per row for any function with context #55281

Describe the bug, including details regarding any error messages, version, and platform.

Gandiva releases the arena after each row only if some function returns a string. has_arena_allocs_ is set only for binary-like return types in BuildFunctionCall.

levenshtein returns int32 but uses arena memory: (shorter_len + 1) * 2 * sizeof(int) bytes per row. In levenshtein(f0, f1) this memory is kept until Evaluate returns. In levenshtein(upper(f0), upper(f1)) it is released after each row.

To reproduce:

1. Make 262,144 rows, two utf8 columns, 16 chars each.
2. Build a projector with levenshtein(f0, f1).
3. Call Evaluate. Compare arrow::default_memory_pool()->total_bytes_allocated() before and after.

Expected: about 1 MB (the output buffers and one arena chunk).
Actual: 36,876,288 bytes. Memory grows with rows x string length. Results are correct. No crash.

Bigger batch, 2,097,152 rows of 64 chars: 1,178 MiB from the arena, peak RSS 1.6 GB.

Update: this also affects external functions. Any function that gets the execution context (kNeedsContext) can take arena memory. If it does not return a string, that memory is kept until the end of the batch. Test: an external C function that copies 64 bytes into the arena and returns int32. 262,144 rows: 17,858,560 bytes from the arena. Expected: about 1 MB.

AI tools helped with the investigation.

Component(s)

C++, Gandiva