perf: Compute array_length of first dimension from offsets

[neilconway]

Which issue does this PR close?

• Closes Compute array_length from offsets #25119.

Rationale for this change

array_length worked by forming a slice of values for each logical row in the given dimension, and then appending the length of that slice to the result. That is inefficient; when computing the length of the first array dimension, we can get that from the offset buffer instead, which is ~200x faster for typical inputs. For other array dimensions, we fallback to the previous code path.

Benchmarks: (M4 Max)

• list/8192, 175.379 µs -> 1.137 µs, -99.35%
• fixed_size_list/8192, 174.770 µs -> 0.963 µs, -99.45%
• list/1, 0.171 µs -> 0.088 µs, -48.48%

What changes are included in this PR?

• Compute array_length of dimension 1 for List and LargeList from offsets
• Compute array_length of dimension 1 for FixedSizeList from the FSL's size
• Treat array_length(x, 1) as equivalent to array_length(x)
• Simplify the code to remove a no-longer-necessary macro
• Add unit tests
• Add benchmark for array_length

What is the testing strategy for this PR?

Existing tests pass; new unit tests added.

Are there any user-facing changes?

No.

[AdamGS]

awesome improvement, LGTM!

[neilconway]

@AdamGS @Rich-T-kid Thanks for the reviews!