fix(external): don't list AIMA's own deployment backend (llama.cpp :8080) as an external service

[rjckkkkk]

Problem

The external-service scanner probes a fixed port list (internal/external/service.go: 8000-8010, 8080, 7860, 5000, 5001, 3000). 8080 is llama.cpp's default port, which an AIMA native deployment also binds. So a model AIMA deployed itself gets surfaced again under "外部服务 / external services" as importable.

Observed on a box where AIMA deployed Qwen3.6-27B-Q4_K_M:

deploy.list   : Qwen3.6-27B-Q4_K_M  runtime=native  address=127.0.0.1:8080  running
external.list : base_url=http://127.0.0.1:8080  source=scan  imported=false  models=["Qwen3.6-27B-Q4_K_M.gguf"]

Same backend, listed in both panels. Worse than cosmetic: clicking import would register a self-referential external-openai backend pointing at AIMA's own deployment.

Fix

Reconciler already holds the proxy, which knows its own (non-external) deployment backends. Exclude any scanned/persisted service whose host:port matches one:

• normalizeHostPort — reduce a base URL/address to a comparable host:port, folding localhost / ::1 / 0.0.0.0 → 127.0.0.1.
• Reconciler.ownDeploymentAddrs — the host:port set of proxy backends where External == false.
• Scan skips own-deployment addresses (no upsert); List filters them out, so a self entry a prior scan already persisted also disappears.

Genuine external services (e.g. Ollama on :11434, LM Studio, a manually-started llama.cpp on a different port) are unaffected.

Tests

• normalizeHostPort table (scheme/path strip, loopback folding).
• List excludes an own :8080 deployment while keeping a genuine Ollama :11434 service (real in-memory DB + proxy with a non-external backend).

go test ./..., go build ./..., go vet, gofmt clean. Verified on real hardware: after the fix, external.list returns the AIMA-deployed :8080 no longer, while deploy.list still shows it running.