Fix updater flavor detection

[i386]

Summary

mesh-llm update now avoids silently replacing a GPU-flavored install with the CPU bundle, and users can explicitly choose or re-detect the best flavor for the current host.

The installer and updater now share the same effective backend preference order:

cuda-blackwell -> cuda -> rocm -> vulkan -> metal -> cpu

Details

• Add command-local mesh-llm update --flavor <flavor> for explicit update bundle selection.
• Add mesh-llm update --detect-flavor to re-run host backend detection and install the best detected release bundle.
• Keep --flavor as the manual update override, and reject combining it with --detect-flavor.
• Detect Blackwell CUDA via nvidia-smi --query-gpu=compute_cap, with NVIDIA /proc model-name fallback.
• Align Linux installer flavor recommendations with updater flavor selection.
• Improve updater probes for Windows ROCm/HIP and Vulkan SDK installs.
• Add comments and docs noting that installer and updater flavor detection must stay in sync.

Validation

• bash -n install.sh
• mocked installer detection for cuda-blackwell, cuda, rocm, vulkan, and cpu
• cargo test -p mesh-llm-system autoupdate
• cargo check -p mesh-llm-system
• cargo check -p mesh-llm-host-runtime
• git diff --check