feat(agent): provider-neutral LlmProvider trait + types (#8)

[ammar-s847]

Implements #8 — the stable seam between Stream A (providers) and Stream C (the reasoning loop).

What

• New core/src/agent/provider/mod.rs: LlmProvider trait (async complete(&ChatRequest, api_key) -> ChatResponse, Send + Sync, object-safe) plus provider-neutral types: ChatMessage/Role, ToolDeclaration, ToolCall, GenerationConfig, ChatRequest (+builders), TokenUsage (+Add), ChatResponse, FinishReason, ProviderError.
• Re-exported from agent::mod.
• API key is passed per call (never stored on the provider) so it can come from the keychain ([E1] Tauri keychain plugin + set/has/delete_provider_key commands #34/[D3] Tauri agent_bridge command + key injection #32) or env fallback ([E4] Env-var fallback for headless/CLI #37) without the provider caring.

MVP / Acceptance (#8)

• ChatRequest carries messages, tool/function declarations, gen config
• ChatResponse carries text, structured tool_calls, token usage
• no reqwest dependency in this module (pure data + trait)
• unit test round-trips a hand-built response (+ object-safety test)

Tests

cargo test -p core agent:: → 21 passed (6 new).

Notes

• ProviderError is intentionally minimal here; [A6] Provider error taxonomy + retry/backoff #13 extends it with retry/backoff classification additively.
• Closes [A1] Define provider-neutral LlmProvider trait + types #8. Unblocks [A2] GeminiProvider over reqwest (generateContent + function calling) #9 (Gemini), [C2] Reasoning loop with bounded tool-call iteration #25 (loop), [C4] Session state (history, usage, steps, cancel token, cortex ctx) #27 (session).