feat: add array_avg scalar function

[crm26]

Summary

Adds array_avg, the last function from the split-PR pipeline tracked in #21536.

Computes the arithmetic mean (sum divided by count) of the elements of a numeric array, returned as Float64. Templated from the merged array_sum (#22542) with the same SQL aggregate NULL conventions; sibling of array_sum, array_product, array_subtract, array_add, array_scale, and array_normalize.

Semantics

NULL semantics — SQL aggregate convention (deliberate divergence from binary-op siblings):

• NULL row → NULL row out
• NULL elements are skipped from BOTH the sum and the count, matching PostgreSQL AVG, DuckDB list_avg, Spark aggregate. So array_avg([1, NULL, 3]) = (1+3) / 2 = 2, not (1+3) / 3.
• All-NULL row → NULL out (matches AVG(...) over an all-NULL column)
• Empty array → NULL (matches sibling array_sum feat: add array_sum scalar function #22542 and array_product Add array_product UDF #22703, PostgreSQL, DuckDB list_avg, SQL Standard AVG-of-empty-set)

Type coercion:

• Inner numeric types (Float32, Int*, UInt*) coerced to Float64. Integer-arg literals are coerced too.
• Return type is always Float64 (since avg of integers can be non-integer).

List shapes supported: List, LargeList, FixedSizeList.

Alias: list_avg (matches the list_sum / list_product pattern).

Test coverage

SLT (array_avg.slt) covers:

• Happy paths (basic, single element, negative values, cancelling positive/negative, non-integer mean [1,2] → 1.5)
• All NULL shapes: bare NULL row, NULL elements skipped, single non-NULL among NULLs, all-NULL array
• Empty array → NULL
• All 3 list shapes (List, LargeList, FixedSizeList)
• Float32, Int64 inner types, integer literals, integer mean that is non-integer
• Multi-row mix
• Error paths (non-list input, zero args, two args)
• Return type assertion (Float64)
• list_avg alias

Closes

The last open slot in #21536 — completes the 8-function split-PR pipeline from the originally-too-big PRs #21371 / #21376.