Use float or higher accumulator for layernorm internally when calculating Mean

[umangyadav]

AMDMIGraphX/src/targets/gpu/kernels/include/migraphx/kernels/layernorm.hpp

Line 61 in 7460ac3

make_array<vec_value_type>(vec_value_type{0}, vec_value_type{0}),

x^2/n can cause accuracy issues when using lower precision i.e. fp16.

Internally it can use higher precision for accumulation.

[pfultz2]

So #2883 will upgrade large reduce_means to use float by default and #2853 will update the x^2/n calculation to be float as well.