Config: unify api_url + lm_proxy_url into a single origin URL + upstream protocol #46

Problem

Today there are two URL knobs pointing at the same upstream:

```ini
[inference]
api_url = http://192.168.0.107:1234/v1 # used by internal inference client (chat, summarize)
lm_proxy_url = http://192.168.0.107:1234 # used by /v1/* pass-through proxy
```

They disagree on whether `/v1` belongs in the URL (see #45), they duplicate the same host:port, and they can silently drift out of sync. The right model is one URL per endpoint, and let Ragger decide what path suffix to tack on based on the upstream protocol it's speaking.

Target config

```ini
[inference]
url = http://192.168.0.107:1234 # origin only — no /v1, no /api/v1
api_key = lmstudio-local
protocol = openai # openai | anthropic (later: lmstudio-stateful)
model = devstral-small-2-24b-instruct-2512@q8_k_xl
```

• One `url` serves both the internal inference client and the LM pass-through proxy. If proxy is enabled, it forwards to the same upstream. No more `lm_proxy_url`.
• `protocol` drives which version-paths Ragger appends:
  • `openai` → `/v1/chat/completions`, `/v1/models`, `/v1/embeddings`
  • `anthropic` → `/v1/messages`
  • future: `lmstudio-stateful` → `/api/v1/chat`, `/api/v1/models` (tied to Proxy: support stateful chat chains (LM Studio native / OpenAI Responses API) #44)
• Version bump (v2) is a protocol concern, not a config concern. User doesn't change `url`.

Downstream (what Ragger exposes to clients)

Always OpenAI-compat for now. OC and other clients hit `/v1/chat/completions`, `/v1/models` on Ragger. Ragger translates to the upstream protocol internally.

Future optional: expose Anthropic-compat `/v1/messages` downstream too, for clients that prefer that shape. Not required for this issue.

Upstream (what Ragger calls)

Driven by `protocol`:

When downstream is OpenAI-compat and upstream is Anthropic, Ragger transforms the request/response shape using the existing format infrastructure. That already exists for the internal inference path (`formats/anthropic.json`) — extending it to proxy means running the same transformer pipeline.

Subsumes / relates to

• Proxy: tolerate trailing /v1 in lm_proxy_url #45 — URL normalization becomes moot because `url` is explicitly origin-only. Error messaging improvements from that issue still apply.
• Proxy: support stateful chat chains (LM Studio native / OpenAI Responses API) #44 — stateful chain support plugs in as another `protocol` value. Clean extension point.

Migration

• On config load, if `api_url` or `lm_proxy_url` is present:
  • Strip any trailing `/v1`, `/v1/`, `/api/v1`, `/api/v1/` to get origin.
  • Log a deprecation notice suggesting the new `url` + `protocol` form.
  • Treat the combined value as `url`. Default `protocol = openai`.
• Remove both old keys in a later release once users have migrated.

Why this matters for the product

The user-facing vision: a single conversation accessible from the Ragger web GUI, Telegram, and any OC-style client, all backed by the same memory store. That requires Ragger to be the sole mediator to the upstream, regardless of which client the user is chatting from. Today's two-URL setup makes it too easy to mis-wire one path and have "Ragger chat" and "OC chat" silently hit different upstreams or miss capture.

Acceptance

• `settings.ini` template + docs show only `url` + `protocol`.
• Old `api_url` / `lm_proxy_url` still accepted with a deprecation warning for at least one release.
• `protocol = anthropic` with appropriate `url` + `api_key` works for both internal inference and the `/v1/*` proxy (OpenAI-shaped requests from OC → Anthropic-shaped upstream calls).
• No `/v1` or `/api/v1` anywhere in user-facing config.