[Bugfix][Kernel]Fix incorrect output tokens when running ChatGLM-9b inferencing on MI250

[jundali2]

[Issue Description]
We encountered random and incorrect output tokens when running ChatGLM-9b inferencing with latest rocm/vllm release on MI250.

[Root Cause]
The vectorized rms_norm_kernel uses the same data type as input tensors (like FP16) to store intermediate results, which causes precision lost and leads to incorrect output tokens.

[Solution]
Use float type to store intermediate results and get correct output tokens.
We compared the outputs from different kernels w/ PyTorch's standard op and verified the final output token correctness, below are the test results:

	Mismatched elements	Greatest absolute difference	Greatest relative difference	Output Tokens	Speed(us)
Non-Vectorized Kernel	1040 / 16777216 (0.0%)	0.000488281	0.001631737	Correct	28.9972
Original Vectorized Kernel	16283524 / 16777216 (97.1%)	0.001953125	0.004055023	In-correct	28.8724
Vectorized Kernel w/ Fix	3840603 / 16777216 (22.9%)	0.000488281	0.000976563	Correct	28.7410

We can see after the fix is applied, without impacting on performance, the "mismatched elements" has reduced significantly, and both the absolute & relative differences are close or even better than non-vectorized kernel, final output tokens are correct as well.