feat: Add Ollama API compatibility layer to proxy

[mmogr]

Description

Add Ollama-native API support to the gglib proxy, making it a drop-in replacement for Ollama on port 11434.

Problem

Apps expecting an Ollama endpoint reject the gglib proxy even when configured on port 11434 because:

• Proxy only served OpenAI /v1/* endpoints
• Ollama clients hit /api/* endpoints (version, tags, chat, generate, embed, etc.) — all returned 404
• Response formats were incompatible:
  • Ollama: NDJSON streaming + flat JSON responses
  • OpenAI: SSE streaming + nested choices array
• Model names were strict (no :latest suffix handling)

Solution

Implement an Adapter pattern with simultaneous dual API surfaces:

• /v1/* — OpenAI-compatible (existing, unchanged)
• /api/* — Ollama-native (new)

Core Features

✅ Ollama Endpoints Implemented:

• GET / — Root probe (returns "Ollama is running")
• GET /api/version — Version info
• GET /api/tags — List models
• POST /api/show — Model metadata
• GET /api/ps — Running models
• POST /api/chat — Chat completions (streaming + non-streaming)
• POST /api/generate — Text generation (streaming + non-streaming)
• POST /api/embed — Embeddings
• POST /api/embeddings — Legacy single-embedding endpoint
• Stubs for /api/pull, /api/delete, /api/copy, /api/create (redirect to CLI)

✅ Format Translation:

• Client request → OpenAI format → llama-server → OpenAI response → Ollama format
• SSE ↔ NDJSON streaming adapter using futures_util::stream::unfold
• Proper timestamp handling via chrono crate

✅ Model Name Normalization:

• Strips :latest suffix automatically (e.g., phi3:latest → phi3)
• Preserves other tags and variants

✅ Code Quality:

• Single unified ProxyState struct shared by both API surfaces
• Extracted helper functions (apply_openai_options, parse_upstream_completion)
• Comprehensive unit tests (10 new tests in ollama_models.rs)
• All 164 existing tests still pass

Files Changed

• New: crates/gglib-proxy/src/ollama_models.rs — Ollama data types + normalization
• New: crates/gglib-proxy/src/ollama_handlers.rs — Route handlers + translation logic
• New: crates/gglib-proxy/src/ollama_stream.rs — SSE→NDJSON streaming adapter
• Modified: crates/gglib-proxy/src/server.rs — Route registration, state unification
• Modified: crates/gglib-proxy/src/lib.rs — Module exports
• Modified: crates/gglib-proxy/Cargo.toml — Added chrono dependency

Testing

• ✅ All 12 proxy unit tests pass
• ✅ All 2 proxy doc-tests pass
• ✅ All 67 runtime tests pass
• ✅ All 34 CLI tests pass
• ✅ All 49 axum tests pass
• Total: 164 tests passing, 0 failures

Usage

# Start proxy on port 11434 (Ollama default)
gglib proxy --port 11434

# Now compatible with any Ollama-expecting client:
curl http://localhost:11434/api/tags
curl -X POST http://localhost:11434/api/chat -d '{"model":"phi3","messages":[...]}'

Breaking Changes

None. Both API surfaces coexist without configuration or toggling.

Closes

TBD (link to related issues if applicable)