[Triton] Add Fused GEMM A8W8 + Split + Concat to Deepseek FP8 Model

[farlukas]

Purpose

Fuse KV projection with torch.cat in DeepSeek R1 0528. torch.cat launches a separate kernel that takes quite some time. We can eliminate this latency by doing the concatenation together with the KV projection in one kernel by using a fused GEMM + split + concat kernel.

This PR depends on ROCm/aiter#1553.

Test Plan

Test Result

---

Essential Elements of an Effective PR Description Checklist

• The purpose of the PR, such as "Fix some issue (link existing issues this PR will resolve)".
• The test plan, such as providing test command.
• The test results, such as pasting the results comparison before and after, or e2e results