docs(server): propose structured inference configuration

[Graffioh]

Summary

• document the current split between CLI parsing, backend preparation,
  backend construction, environment variables, ServerConfig, and model cards
• propose BackendArgs -> BackendPlan -> create_backend(plan) as the first
  implementation boundary
• keep observed runtime facts in a later BackendInfo report and serving
  concerns in ServerConfig
• stage the argument registry, config-file loader, environment aliases, and
  runtime reporting as follow-up work
• compare the proposal with current vLLM, SGLang, TensorRT-LLM, TGI, and
  llama.cpp configuration designs

Scope

This is a documentation-only RFC. It does not change runtime behavior or add a
configuration file.

Why this order

The factory currently receives both raw BackendArgs and a partial resolved
plan. server_main can mutate DDTree and SpecLA fields after preparation, so
the factory must re-check the pair. Sealing construction behind one immutable
plan removes that invalid state before Lucebox adds another configuration
source.

Verification

• based on Luce-Org/lucebox@298031aa
• checked the described symbols and call sites against current source
• verified all referenced local paths and external design links
• git diff --cached --check
• confirmed balanced Markdown code fences