fix(cuda): sanitize invalid Blackwell smpbo values

[peter941221]

Fixes #23385.

Failure mechanism:

• early Blackwell launch drivers can report broken cudaDeviceProp.sharedMemPerBlockOptin values such as 0 or 0x100000001
• ggml_cuda_init() stored that value directly in info.devices[id].smpbo
• multiple CUDA paths consume smpbo as a hard shared-memory ceiling, not just softmax
• mmq.cuh uses it to prune valid MMQ tile shapes, which can leave mmq_x_best = 0 and abort on Blackwell
• mmid.cu and softmax.cu also use the same cached limit when setting dynamic shared memory attributes

Semantic change:

• sanitize sharedMemPerBlockOptin once during CUDA device-info initialization
• if the reported value is 0 or implausibly large (> 1 MiB), log a warning and fall back to sharedMemPerBlock
• keep the workaround centralized so every smpbo consumer gets the same sane value

Why here instead of another per-kernel workaround:

• PR fix Blackwell sharedMemPerBlockOptin overflow causing SOFT_MAX error #22338 already demonstrated the same Blackwell driver bug on a softmax path
• issue Eval bug: Fatal MMQ crashes on Blackwell (RTX 5090/5080) due to unhandled sharedMemPerBlockOptin driver bug #23385 shows the remaining MMQ failure comes from the same cached device property
• fixing the cached property is the narrowest place that covers MMQ, MMID, and SOFT_MAX together

Validation:

• source-path verification:
  • ggml/src/ggml-cuda/ggml-cuda.cu initializes info.devices[id].smpbo
  • ggml/src/ggml-cuda/mmq.cuh gates MMQ tile selection on smpbo
  • ggml/src/ggml-cuda/mmid.cu and ggml/src/ggml-cuda/softmax.cu also consume smpbo
• local Windows configure passed with -DGGML_CUDA=ON -DCMAKE_CUDA_ARCHITECTURES=120
• targeted build did not finish within the local command timeout window, so runtime reproduction is still pending on the patched tree

[ggml-gh-bot]

Hi @peter941221, thanks for your contribution!

Per our contribution guidelines, the automated PR checker found the following issue(s) that need your attention:

• Multiple open PRs from a new contributor: We limit new contributors (those without a previously merged PR) to 1 open PR at a time. You currently have 3 open PRs.

---

Please note that maintainers reserve the right to make final decisions on PRs. If you believe there is a mistake, please comment below.