float16 arrays: np.mean upcasts, np.var/np.std/np.sum do not and hit inf

NumPy 2.5.4, uniform(0,1) data stored as float16 (max 65504). np.mean is safe: _mean in numpy/_core/_methods.py switches the accumulator to float32 for float16 input (1e6 values: 0.5, relative error 3e-4). _var has no such branch, so x.var() returned inf from 2e5 values on (x.std() too, tested at 1e6), with the overflow warning raised in the float16 sum inside _var; the true variance was 0.083. x.sum() also returned inf from 2e5 values and was only about 2e-4 accurate below that. np.cumsum was useless (59% low at 1e4 values). Pass dtype=np.float32 to sum/var/std: var(dtype=np.float32) gave 0.08340139 vs 0.08340138 true, and sum(dtype=np.float32) was within 2e-7.