[Bugfix] hotfix for gptq-marlin non-contiguous error

[DefTruth]

fix #14887 #15386

Hotfix for gptq-marlin non-contiguous error. Previous #15319 only fixed the GPTQ code path and #14946 only fixed the AWQ code path. We should lower the non-contiguous fix into the ops.gptq_marlin_gemm function, since gptq_marlin_gemm has been used in multiple code paths, both AWQ and GPTQ.

# GPTQ
MLA -> gptq_marlin.py:341 -> mixed_precision/marlin.py:123 (previous fixed here) ->  marlin_utils.py:334 apply_gptq_marlin_linear -> _custom_ops.py:741 ops.gptq_marlin_gemm -> torch.ops._C.gptq_marlin_gemm
# AWQ
MLA -> awq_marlin.py:303 -> marlin_utils.py:379 apply_awq_marlin_linear -> _custom_ops.py:741 ops.gptq_marlin_gemm -> torch.ops._C.gptq_marlin_gemm

Performance w/ or w/o prefix cache for R1-AWQ model: (TTFT from 13701.38ms -> 1183.58ms, this PR don't downgrade the performance of TTFT)

# MLA + AWQ + w/o prefix cache + NVIDIA L20 + pp2 + tp8
python3 benchmark_serving.py \
        --backend vllm \
        --model /workspace/DeepSeek-R1-awq \
        --port 8862 \
        --endpoint /v1/completions \
        --dataset-name random \
        --dataset-path ${SHAREGPT_DATASET_PATH} \
        --random-prefix-len 0 \
        --random-input-len 4096 \
        --random-output-len 1024 \
        --ignore-eos \
        --max-concurrency 32 \
        --num-prompts 32

Maximum request concurrency: 32
============ Serving Benchmark Result ============
Successful requests:                     32
Benchmark duration (s):                  130.67
Total input tokens:                      131072
Total generated tokens:                  32768
Request throughput (req/s):              0.24
Output token throughput (tok/s):         250.77
Total Token throughput (tok/s):          1253.85
---------------Time to First Token----------------
Mean TTFT (ms):                          13701.38
Median TTFT (ms):                        14286.95
P99 TTFT (ms):                           24786.87
-----Time per Output Token (excl. 1st token)------
Mean TPOT (ms):                          113.26
Median TPOT (ms):                        112.76
P99 TPOT (ms):                           124.87
---------------Inter-token Latency----------------
Mean ITL (ms):                           113.26
Median ITL (ms):                         104.52
P99 ITL (ms):                            853.17
==================================================

# MLA + AWQ + w/ prefix cache + NVIDIA L20 + pp2 + tp8
python3 benchmark_serving.py \
        --backend vllm \
        --model /workspace/DeepSeek-R1-awq \
        --port 8862 \
        --endpoint /v1/completions \
        --dataset-name random \
        --dataset-path ${SHAREGPT_DATASET_PATH} \
        --random-prefix-len 3072 \
        --random-input-len 1024 \
        --random-output-len 1024 \
        --ignore-eos \
        --max-concurrency 32 \
        --num-prompts 32

## TTFT from 13701.38ms -> 1183.58ms
Maximum request concurrency: 32
============ Serving Benchmark Result ============
Successful requests:                     32
Benchmark duration (s):                  112.73
Total input tokens:                      131072
Total generated tokens:                  32768
Request throughput (req/s):              0.28
Output token throughput (tok/s):         290.69
Total Token throughput (tok/s):          1453.45
---------------Time to First Token----------------
Mean TTFT (ms):                          1183.58
Median TTFT (ms):                        1198.95
P99 TTFT (ms):                           1573.84
-----Time per Output Token (excl. 1st token)------
Mean TPOT (ms):                          108.95
Median TPOT (ms):                        108.94
P99 TPOT (ms):                           109.46
---------------Inter-token Latency----------------
Mean ITL (ms):                           108.95
Median ITL (ms):                         109.81
P99 ITL (ms):                            120.19
==================================================

This PR only acts as a temporary workaround. For performance reasons, the best approach is to make gptq_marlin_gemm support non-contiguous input in the future or carefully manage context_chunk_workspace for MLA (don't fuse nope and pe kv cache workspace into a single Tensor).

@LucasWilkinson @mgoin @robertgshaw2-redhat PTAL ~

[github-actions]

👋 Hi! Thank you for contributing to the vLLM project.

💬 Join our developer Slack at https://slack.vllm.ai to discuss your PR in #pr-reviews, coordinate on features in #feat- channels, or join special interest groups in #sig- channels.

Just a reminder: PRs would not trigger full CI run by default. Instead, it would only run fastcheck CI which starts running only a small and essential subset of CI tests to quickly catch errors. You can run other CI tests on top of those by going to your fastcheck build on Buildkite UI (linked in the PR checks section) and unblock them. If you do not have permission to unblock, ping simon-mo or khluu to add you in our Buildkite org.

Once the PR is approved and ready to go, your PR reviewer(s) can run CI to test the changes comprehensively before merging.

To run CI, PR reviewers can either: Add ready label to the PR or enable auto-merge.

🚀

[DefTruth]

Close, since #14658 has been merged.