Proxy: tolerate trailing /v1 in lm_proxy_url #45

Problem

`lm_proxy_url` is documented/felt to be an OpenAI-style base URL (`http://host:port/v1\`), but the proxy code prepends `/v1/*` itself. When users follow the `/v1` convention, URLs become `http://host:port/v1/v1/models\` and curl returns "Proxy connection failed" with nothing actionable in the log.

Affected call sites in `src/inference.cpp`:

• `proxy_list_models()` → `proxy_request("/v1/models", ...)`
• chat forwarding passes `/v1/chat/completions` through unmodified
• streaming proxy same

Fix options

1. Normalize on parse. In `config.cpp`, strip a trailing `/v1` (with or without slash) from `lm_proxy_url` before storing. Simple, invisible, matches user expectation.
2. Normalize on use. In `forward_request`, if `base` ends in `/v1`, skip the `/v1` prefix in `path`. Slightly more magic, but handles CLI overrides too.
3. Document and leave strict. Update `settings.ini` template + docs to be explicit that `lm_proxy_url` is origin-only, no `/v1`.

Preferred: (1) + better error messaging — strip the suffix during parse, and if `curl_easy_perform` fails, include the actual `curl_easy_strerror(res)` string in the error log rather than the generic "Proxy connection failed." Latter helps diagnose any remaining issue (DNS, TLS, refused) without a rebuild.

Acceptance

• Setting `lm_proxy_url = http://host:1234/v1\` works identically to `http://host:1234\`.
• Log lines on proxy failure include the underlying curl error string.