Surface CUTLASS initialization and CUDA launch-configuration failures

Problem

Two initialization boundaries currently discard errors:

1. setup.py runs git submodule update --init cutlass without checking its exit status and without setting the working directory to the repository root. If Git fails, installation continues until compilation reports a less actionable missing-header error.
2. The forward launcher discards the return values of both cudaFuncSetAttribute calls used to configure dynamic shared memory.

The documented clone and submodule workflow itself works when initialization succeeds. The problem is that failures are reported later and at the wrong boundary.

Desired outcome

• Run CUTLASS initialization from the repository root.
• Stop immediately on a missing Git executable or nonzero Git exit.
• Include enough context in the error for users to retry the documented submodule command.
• Propagate cudaFuncSetAttribute errors through the project's PyTorch/CUDA error mechanism.
• Preserve successful setup behavior, architecture selection, launch dimensions, and shared-memory sizes.

Start with setup.py and locate the forward launcher’s two cudaFuncSetAttribute calls. Check how the documented submodule command is run and how PyTorch/CUDA errors are propagated, then verify that Git failures and launch-configuration failures stop at their respective boundaries while successful setup, architecture selection, launch dimensions, and shared-memory sizes remain unchanged.