[Feat][Plugin] Enable DeepSeek-V3.2 for vLLM OOT Plugin

[kliuae-amd]

Motivation

This PR is a follow up to #399 in adding DeepSeek-V3.2 support to ATOM's vLLM plugin mode.
Currently this PR contain changes from #399 and should be much more simplified after it gets upstreamed.

Technical Details

Test Plan

Accuracy test with lm_eval

Model: deepseek-ai/DeepSeek-V3.2

Server command:

ATOM_DISABLE_VLLM_PLUGIN=0 \
ATOM_DISABLE_VLLM_PLUGIN_ATTENTION=0 \
vllm serve deepseek-ai/DeepSeek-V3.2 \
  -tp 8 \
  --gpu-memory-utilization 0.8 \
  --no-enable-prefix-caching \
  --disable-uvicorn-access-log \
  --trust-remote-code \
  --compilation-config '{"cudagraph_mode": "FULL_AND_PIECEWISE"}' \
  --kv-cache-dtype {auto, fp8} \
  --block-size 1

Test Result

lm_eval command

lm_eval --model local-completions  --model_args model=deepseek-ai/DeepSeek-V3.2,base_url=http://localhost:8000/v1/completions --batch_size 100  --tasks gsm8k --num_fewshot 20

vLLM plugin, bf16 kv cache

Tasks	Version	Filter	n-shot	Metric		Value		Stderr
gsm8k	3	flexible-extract	20	exact_match	_	0.9545	_	0.0057
		strict-match	20	exact_match	_	0.9545	_	0.0057

vLLM plugin, fp8 kv cache

Tasks	Version	Filter	n-shot	Metric		Value		Stderr
gsm8k	3	flexible-extract	20	exact_match	_	0.9431	_	0.0064
		strict-match	20	exact_match	_	0.9393	_	0.0066

Performance on MI355X, TP8

ISL/OSL	Concurrency	KV Cache	vLLM Plugin Req/s	ATOM Req/s	vLLM Plugin over ATOM (Req/s)	vLLM Plugin Total tok/s	ATOM Total tok/s	vLLM Plugin over ATOM (tok/s)
1k/1k	128	fp8	3.51	3.80	-7.63%	7194.55	7786.85	-7.61%
1k/1k	128	bf16	3.47	3.58	-3.07%	7099.09	7326.96	-3.11%
1k/1k	64	fp8	2.28	2.32	-1.72%	4677.17	4744.63	-1.42%
1k/1k	64	bf16	2.26	2.22	+1.80%	4621.56	4538.04	+1.84%
8k/1k	128	fp8	1.42	1.59	-10.69%	13077.77	14614.80	-10.52%
8k/1k	128	bf16	1.39	1.03	+34.95%	12831.04	9505.53	+34.99%
8k/1k	64	fp8	1.16	1.22	-4.92%	10683.58	11288.62	-5.36%
8k/1k	64	bf16	1.16	0.86	+34.88%	10731.82	7962.13	+34.79%

Submission Checklist

• Look over the contributing guidelines at https://github.com/ROCm/ROCm/blob/develop/CONTRIBUTING.md#pull-requests.