feat(backend): add llama.cpp (llama-server) backend wrapper

[bilersan]

Contributes to #92

Summary

Adds llamacpp as a named backend alongside vllm, openai, anthropic, ollama, and lmstudio. Follows the established vLLM pattern exactly:

• llamacpp struct embedding *openAICompat with cold-start retry on ECONNREFUSED
• Ping() delegates to coldStartRetry (reuses vllm_internal.go helper — no duplication)
• Default endpoint: http://localhost:8080 (llama-server default)
• No API key required by default (llama-server typically runs unauthenticated)

.ctxrc usage (once factory wiring + consumer commands land)

backends:
  - name: local
    type: llamacpp
    endpoint: http://localhost:8080
    timeout: 60s
default_backend: local

What this PR does NOT include

This PR adds the backend type and tests only. It intentionally does not include:

• Factory registration wiring — not yet implemented for any backend in this branch
• Consumer CLI commands (ctx ai, ctx compact --emit, ctx ingest) — expected in a future phase

When the factory wiring and consumer commands land, llamacpp will be available as a backend type with zero additional work.

Why llama.cpp needs its own wrapper (vs generic openai-compatible)

llama-server behaves like vLLM during model loading: the TCP listener is not yet bound while weights load, so the OS returns ECONNREFUSED at the socket level (not HTTP 503). The coldStartRetry logic from vllm_internal.go handles exactly this case. A generic openai-compatible backend would fail immediately on connection refused instead of retrying.

Validation

Tested against a live llama-server running Qwen3-4B-Q4_K_M:

Test	Result
Ping → GET /v1/models	HTTP 200, model listed
Complete → POST /v1/chat/completions	Model response received
Cold-start retry (unit)	Retries ECONNREFUSED, stops on non-dial errors
Non-dial error (unit)	HTTP 500 returns immediately, no retry
All existing backend tests	PASS (no regressions)

Files changed

File	Change
internal/config/backend/backend.go	+2 constants (NameLlamaCpp, DefaultEndpointLlamaCpp)
internal/backend/types.go	+1 struct (llamacpp)
internal/backend/llamacpp.go	Constructor + Ping override (~60 lines)
internal/backend/llamacpp_test.go	3 unit tests (mock httptest server)
internal/backend/llamacpp_e2e_test.go	2 e2e tests (build tag: e2e, requires live server)