Multinomial FP8 verification test is failing

[umangyadav]

#### Ref ######
Run instruction: @6 = ref::exp(@5) -> fp8e4m3fnuz_type, {2, 5}, {5, 1}, target_id=0
Output: 0.9375, 1, 0.5625, 0.75, 0.8125, 0.5625, 1, 0.6875, 0.9375, 0.5625
Run instruction: @7 = ref::prefix_scan_sum[axis=1,exclusive=0,reverse=0](@6) -> fp8e4m3fnuz_type, {2, 5}, {5, 1}, target_id=0
Output: 0.9375, 2, 2.5, 3.25, 4, 0.5625, 1.5, 2.25, 3.25, 3.75


##### GPU ######
Run instruction: @5 = gpu::code_object[code_object=10224,symbol_name=reduce_max_sub_exp_convert_kernel,global=128,local=64,](input,@4) -> float_type, {2, 5}, {5, 1}, target_id=0
Output: 0.9375, 1, 0.5625, 0.75, 0.8125, 0.5625, 1, 0.6875, 0.9375, 0.5625
Run instruction: @7 = gpu::prefix_scan_sum[axis=1,exclusive=0,reverse=0](@5,@6) -> float_type, {2, 5}, {5, 1}, target_id=0
Output: 0.9375, 1.9375, 2.5, 3.25, 4.0625, 0.5625, 1.5625, 2.25, 3.1875, 3.75

#2510 adds FP8 test for the multinomial op which is failing because prefix_scan_sum is producing different results for the "ref" and "gpu" target.

Numbers are close for both target, but Float allows for higher precision and error accumulates through prefix_scan_sum.