[Gluon] [Triton] [MI450] [MI350] Enable Unified Attention option for decode

[k50112113]

This PR enables Unified Attention option for decode (Triton for MI350 and Gluon for MI450)

On MI350, the implementation here is just to verify the results on MI350. The primary purpose of this PR is to enable Gluon Unified Attention for decode on MI450.

Triton/Gluon Unified Attention right now supports shuffling of both Key and Value cache, therefore, I also edited atom/model_engine/model_runner.py to fix the layout for Value cache.

I added an env var ATOM_ENABLE_TRITON_UNIFIED_ATTENTION_DECODE for the user to toggle between Unified Attention or Gluon Paged Attention.

The block size has to be fixed to 64 for BF16 KV cache and 128 for FP8 KV cache, so I also added the switch at atom/model_ops/attentions/aiter_attention.py

This PR depends on ROCm/aiter#2472

Server commend:

model_path="/data/openai/gpt-oss-120b"
export ATOM_ENABLE_TRITON_UNIFIED_ATTENTION_DECODE=1
python -m atom.entrypoints.openai_server \
  --model $model_path --kv_cache_dtype fp8 --block-size 1024 -tp 1 --torch-profiler-dir /app/_test/trace --mark-trace

lm_eval results (TP1):

local-completions ({'model': '/data/openai/gpt-oss-120b', 'base_url': 'http://localhost:8000/v1/completions', 'num_concurrent': 64, 'max_retries': 1, 'tokenized_requests': False}), gen_kwargs: ({}), limit: None, num_fewshot: 3, batch_size: 1
|Tasks|Version|     Filter     |n-shot|  Metric   |   |Value |   |Stderr|
|-----|------:|----------------|-----:|-----------|---|-----:|---|-----:|
|gsm8k|      3|flexible-extract|     3|exact_match|↑  |0.4390|±  |0.0137|
|     |       |strict-match    |     3|exact_match|↑  |0.2259|±  |0.0115|

lm_eval results (TP8)

local-completions ({'model': '/data/openai/gpt-oss-120b', 'base_url': 'http://localhost:8000/v1/completions', 'num_concurrent': 64, 'max_retries': 1, 'tokenized_requests': False}), gen_kwargs: ({}), limit: None, num_fewshot: 3, batch_size: 1
|Tasks|Version|     Filter     |n-shot|  Metric   |   |Value |   |Stderr|
|-----|------:|----------------|-----:|-----------|---|-----:|---|-----:|
|gsm8k|      3|flexible-extract|     3|exact_match|↑  |0.4405|±  |0.0137|
|     |       |strict-match    |     3|exact_match|↑  |0.2153|±  |0.0113|