remove compiled_rmsnorm

[tianyu-l]

Stack from ghstack (oldest at bottom):

• -> remove compiled_rmsnorm #535

It seems this feature (#442) has been very unreliable, e.g.

• it did not work well with TP in terms of numerics, despite passing unit test (add unit test for compiled_rmsnorm #506)
• it does not work with PP as expected (still having the cuda context error which fused_rmsnorm experienced)
• it used to work in 1D but seems to have no effect today (08/19), not sure why

Let's remove it for now. To get performant RMSNorm:

• apply (block-level) torch.compile, or
• use fused_rmsnorm in 1D or 2D (which also doesn't support PP well)
• wait for a dedicated CUDA kernel

This PR also adjusts integration tests as we now use rmsnorm as default.