feat: add a repeatable PARAM keyword for engine flags

[outofcoffee]

Spinloop files can now carry their engine's flags with repeatable PARAM lines, in place of a preset file.

Summary

• New repeatable PARAM key = value instruction in Spinloop files; keys are read in the engine's own flag dialect, so ngl becomes --n-gpu-layers for llama.cpp while vLLM's and MTPLX's long-form keys pass through unchanged
• Precedence where the same flag is set in more than one place: the Spinloop's typed instructions, then a PARAM, then the preset
• A PARAM may name the model where the engine takes it as a flag — llama.cpp's hf/model/hf-file, MTPLX's model — so a Spinloop with no preset file still serves and deploys its model; vLLM takes its model positionally, so a PARAM cannot stand in for MODEL there
• spinloop cloud deploy and a fleet node wake read PARAMs the way they read a preset, dropping the settings the destination owns from them alike
• Docs updated (the Spinloop file, serve, cloud) and two new examples: examples/llamacpp/qwen3.8-27b-mtp, PARAM-based with no preset file, and examples/vllm/qwen3.8-27b, the first vLLM example

Implementation details

• internal/spinloop gains the Param pair and its parse/format; internal/preset gains Dialect.Merge(layers ...) so the PARAM layer is one more layer in the same merge a preset already goes through — the dialect rendering, the owned-flag dropping, and the model-reference handling all fall out of that one change
• Model sourcing is defined once per engine: modelSourceFlags(runner) lists the flags that name the model (llamacpp hf-repo/model/hf-file, mtplx model, none for vllm), and paramModel is the shared core both the serve and the deploy paths call