[CK_TILE] Improve RMS/Layer Normalization 2 Pass Pipeline Performance

[ruanjm]

Improve RMS Normalization Performance
Mainly forcus on tile shape 64 x 65536.

Note

Current setting may not be optimal for other tensor shapes especially for the cases that m is not 64.

1. Set RepeatN as 1 for vectorized-8 cases.
2. Fine tuned thread grid size for various vectorized size under 64 x * cases. Note that performance tests are based on aiter test https://github.com/ROCm/aiter/blob/main/op_tests/test_rmsnorm2d.py. Smoke tests in CK (https://github.com/ROCm/composable_kernel/tree/develop/example/ck_tile/10_rmsnorm2d/script) is only used for reference.
3. In fused-add-store cases, use residual buffer in 2nd pass for saving memory transactions and calculations.

[spolifroni-amd]

Nothing to review for docs.