rms norm: improve fwd perf

[v0i0]

about doubles perf for B=32K, H=2K, presumably before was storing a bunch into inv_sqrt?
might also indicate something funny happening with keepdim.

[v0i0]

perf results on h100:

32768,256,fwd,helion_orig,1642.1788056405082
32768,256,fwd,helion_new,1578.5740566290297
32768,512,fwd,helion_orig,1817.2547849940113
32768,512,fwd,helion_new,1866.2814686329166
32768,1024,fwd,helion_orig,1488.1836334821526
32768,1024,fwd,helion_new,2016.8193458851026
32768,2048,fwd,helion_orig,935.4300356703114
32768,2048,fwd,helion_new,2087.6677455652543
32768,4096,fwd,helion_orig,242.58288588525073
32768,4096,fwd,helion_new,2128.879800252783
32768,8192,fwd,helion_orig,187.45590805263205
32768,8192,fwd,helion_new,1800.378577328567
32768,16384,fwd,helion_orig,181.87879046800123
32768,16384,fwd,helion_new,232.25287573963158
32768,32768,fwd,helion_orig,188.59863561089344
32768,32768,fwd,helion_new,201.6228783951727