CLI: clarify backend selection (no --backend flag; CUDA not wired to LLM path)

[pekkah]

Summary

Users may try --backend cuda based on the README mentioning a CUDA backend, but no such flag exists on the RunCommand. The flag is silently ignored by Spectre.Console and inference falls back to CPU. The CUDA backend (SharpInference.Cuda) is wired only to ImageCommand (image generation), not to LLM inference.

What exists

• -g 0 (default): CPU only
• -g N: hybrid - N layers on Vulkan GPU + rest on CPU
• -g -1: all layers on Vulkan GPU (auto-detect via TierPlanner)

Vulkan only. Requesting CUDA for chat inference is not supported.

Suggested fixes

1. Reject unknown options (Spectre.Console has StrictParsing mode) instead of silently ignoring --backend cuda
2. Add a --backend {cpu,vulkan} option that maps onto -g semantics, and emit an explicit error if cuda is requested for chat inference
3. README cleanup: clarify that CUDA is image-only

[pekkah]

Resolved by 5473e03 (Add CUDA text-inference backend and close the perf gap vs Vulkan): the --backend flag exists (src/SharpInference.Cli/RunCommand.cs:112, values auto|vulkan|cuda) and CUDA is wired all the way into the LLM forward pass via CudaForwardPass. Verified just now with --backend cuda -g -1 against Qwen3-8B-Q4_K_M; CUDA backend activates, prefill 64.7 t/s.