fix(agg): use f64 sum for integer AVG to avoid overflow

[ser-vasilich]

Summary

The scalar reduction path computed AVG on integer columns as
(double)acc.sum_i / cnt where sum_i is accumulated through
unsigned wrap (to dodge UBSan). For columns whose true sum exceeds
2^63 — e.g. an i64 user-id column with signed values spread across
±9e18 over millions of rows, true sum ~7.6e24 — the wrap leaves
garbage in sum_i and the mean is wrong by orders of magnitude.

Add a parallel f64 sum_d field to reduce_acc_t, populate it in
every integer reduction kernel (REDUCE_LOOP_I, BOOL/U8 path) and in
reduce_merge. OP_AVG (and VAR/STDDEV mean) now read sum_d
instead of sum_i. F64 path is unchanged.

Reduction kernels gain one f64 add per element; vectorises with the
existing integer SIMD store.