[None][perf] enable paged context for skip-softmax fp8 kv

[bobboli]

Summary

• Force paged-context FMHA when skip-softmax attention is used with FP8 KV cache.
• This keeps context QKV on the paged-context preprocessing path so FP8 KV-cache skip-softmax dispatch does not fall back to BF16 packed context.

Testing

• python3 -m py_compile tensorrt_llm/_torch/attention_backend/trtllm.py
• git diff --check HEAD^ HEAD

Not run

• pytest/runtime verification in tekit4. Local Python is missing pluggy/torch, and I did not build TensorRT-LLM in this checkout.
• Full-file ruff on trtllm.py reports pre-existing style violations unrelated to this change.