Add HyperSphere Normalization

[gkielian]

Test out norm variation that simply maps to the surface of a hypersphere.

[gkielian]

Add hypersphere norm variation #361

Relates to the above pull request

[gkielian]

Results from sweep on small model definitely show some affects of radius, though it is indeed around the sqrt embedding dim:

May need to test on a larger network in order to see more definitive differences from slight effects like removing the gain, however it seems for this network hyperspherenorm is really just as good as rmsnorm for conventional radius.