[kernel] feat: vectorized kernels

[AlpinDale]

So far, RMSNorm is done. No decrease in accuracy:

Before:

|Tasks|Version|     Filter     |n-shot|  Metric   |   |Value |   |Stderr|
|-----|------:|----------------|-----:|-----------|---|-----:|---|-----:|
|gsm8k|      3|flexible-extract|     5|exact_match|↑  |0.6808|±  |0.0128|
|     |       |strict-match    |     5|exact_match|↑  |0.6740|±  |0.0129|

After:

|Tasks|Version|     Filter     |n-shot|  Metric   |   |Value |   |Stderr|
|-----|------:|----------------|-----:|-----------|---|-----:|---|-----:|
|gsm8k|      3|flexible-extract|     5|exact_match|↑  |0.6816|±  |0.0128|
|     |       |strict-match    |     5|exact_match|↑  |0.6748|±  |0.0129|

Performance (Olmoe-1B-7B):

• RMSNorm
• Activations
• RoPE