fix(dflash): auto-detect GPU arch to prevent sm_120a on consumer Blackwell

[easel]

Problem

On CUDA 13.2+ with consumer Blackwell GPUs (for example, RTX 5090, SM 12.0), using an unset
CMAKE_CUDA_ARCHITECTURES or native can resolve to sm_120a instead of
sm_120, which can trigger CUDA_ERROR_ILLEGAL_INSTRUCTION at runtime on
consumer hardware.

Fix (auto-detect only)

• At configure time, if CMAKE_CUDA_ARCHITECTURES is unset or native, run
  nvidia-smi --query-gpu=compute_cap --format=csv,noheader.
• Parse the compute capability (for example 12.0) and set
  CMAKE_CUDA_ARCHITECTURES explicitly (for example 120).
• Keep the change isolated to dflash/CMakeLists.txt so it can be reviewed and
  merged independently from consumer-specific workaround behavior.

Test plan

• cmake -B build -S dflash/ prints dflash27b: GPU compute_cap 12.0 → CUDA_ARCHITECTURES=120 on Blackwell hardware.
• cmake --build build succeeds without CUDA-arch related compiler/runtime errors on a Blackwell consumer system.

[davide221]

@easel thanks for the contribution! Is the speed problem still present ?

[easel]

@easel thanks for the contribution! Is the speed problem still present ?

Yes. I think it's related to the workflow -- I'm putting together a small benchmark script to compare.

[easel]

This may not be necessary if expectation is to always build multi-arch binary. I ran into it because claude got excited about optimizing and ended up with a slightly incompatible build.