[Quantization] [Performance] Enable Marlin GEMM kernels for the calibration-free RTN-based quantization

[sakogan]

This PR enhances the work started in #18768 and #20766 by enabling Marlin kernels for the calibration-free RTN-based quantization.

These kernels substantially improve the performance of dense/MoE models quantized with RTN.

We ran the built-in latency benchmark with several Llama models on a machine equipped with H100 GPUs. The exact command was
[RTN_NUM_BITS=4] vllm bench latency --model <model> --n 1 --num-iters-warmup 3 --num-iters 10 --input-len 256 --output-len 32 -tp <#GPUs> --batch-size <batch> -q rtn
Each data point is an average of 5 runs, the units are seconds (measuring generation latency, the lower the better).

Here are the results for Llama3.1-8B (ran on 1 GPU), for various batch sizes (old/new refer to pre-PR/post-PR implementations):

Variant (data type)	1	4	8	16
old RTN (Int8)	0.458	0.490	0.516	0.570
new RTN (Int8)	0.174	0.222	0.285	0.402
old RTN (Int4)	0.720	0.761	0.794	0.847
new RTN (Int4)	0.139	0.180	0.230	0.331

Here are the results for Llama3.3-70B (ran on 4 GPUs), for various batch sizes:

Variant (data type)	1	4	8	16
old RTN (Int8)	1.104	1.190	1.262	1.411
new RTN (Int8)	0.416	0.545	0.707	1.025
old RTN (Int4)	1.736	1.859	1.933	2.082
new RTN (Int4)	0.328	0.442	0.573	0.846

[mergify]

This pull request has merge conflicts that must be resolved before it can be
merged. Please rebase the PR, @sakogan.

https://docs.github.com/en/pull-requests/collaborating-with-pull-requests/working-with-forks/syncing-a-fork

[mgoin]

Looks reasonable to me, thanks for using the existing kernel! Please fix the merge conflict and I will enable the full CI @sakogan

[sakogan]

Looks reasonable to me, thanks for using the existing kernel! Please fix the merge conflict and I will enable the full CI @sakogan

@mgoin Done (and thanks for the review!)