docs: use Grouped GEMM for the agent launch walkthrough

[YiyanZhai]

The launch walkthrough currently uses KDA. Use Grouped GEMM for the introductory task, while retaining the recorded KDA examples in the following diagnostic and profiling chapters.

• Use the harness's grouped_gemm_fp8 task and unified setup/evaluation commands, covering four configurations with DeepGEMM as the correctness reference and performance baseline.
• Keep the task introduction short, have the agent implement the TIRx kernel, and explain the transition to KDA in subsequent chapters.
• Use the standard harness dependency installation for DeepGEMM and document frontier results using the new repository's existing Launching the Agent chapter.

Validation: strict HTML build (make html, Sphinx warnings treated as errors) and git diff --check passed. GPU end-to-end validation remains pending.

Dependency: TIRx-harness #1 migrates the four-configuration task to the new Harness repository and adds the rebuilt DeepGEMM wheels to its benchmark/server dependencies. Merge that PR before using the documented default-branch setup.

Harness now has the packaging changes in its linked PR, with wheels for Python 3.12 and 3.13. The chapter states the wheel runtime requirements: Linux x86_64, glibc >= 2.38, and CUDA Toolkit 13.2.