This was described in the RMSNorm paper as being able to accomplish the
same task as RMSNorm usually with only performing calculations on the
first 6% of entries.

This is because the average of the RMSNorm changes more slowly the more
items are added, and the RMSNorm authors noted that the tokens they
measured had around the same value.

Add variation of pRMSNorm that simply sets a max number of tokens to
utilize for the RMSNorm normalization.

This should be much easier to implement in hardware.