docs(readme): lead with a sequential setup and make BYOK the first decision

The quick start used to be a `cp .env.example .env` followed, fifty lines
later, by a separate "Bring your own LLM" section. That ordering buried the
one choice the project cannot make for the reader: which model to point it at.

Restructured as five numbered steps -- prerequisites, get the code, choose your
model, add embeddings, install and log in -- with the provider recipes as
collapsible blocks inside step 3, so a reader sees the one that applies rather
than scrolling past four that do not. Reference material (ports, config table,
telemetry, MCP, development) moves below the fold; it is day-30 content, not
day-0.

Corrections while there:

* "there is no Maven wrapper in the repo" is no longer true -- ./mvnw is
  committed and pins the version CI uses.
* `git clone <this-repo>` was still a placeholder.
* The prerequisites omitted git and buildx, and a vanilla Amazon Linux 2023
  image has neither, so step 2 failed before DeepSQL was involved.
* Promoted two facts that were buried but bite people: DEEPSQL_CHAT_PROVIDER
  gating its whole group, and the 3072-dimension embedding requirement.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>

Author: geekypunk