feat(agent): add AI Agent conversation APIs

[sunli829]

Summary

• New agent package (AgentContext) covering the AI Agent conversation API added in longbridge/openapi#557 (docs: longbridge/developers#1129):

  • Workspaces — GET /v1/ai/workspaces
  • Agents — GET /v1/ai/workspaces/{id}/agents
  • Conversation / Continue — blocking POST .../conversations and .../continue
  • ConversationStream / ContinueStream — same endpoints with Accept: text/event-stream, returning a *ConversationStream iterator (Next/Event/Err/Close, bufio.Scanner-style)
  • ConversationStreamEvent is modeled as an interface with 7 concrete event types (ChatStartedEvent, WorkflowStartedEvent, MessageEvent, PingEvent, ChatFinishedEvent, WorkflowFinishedEvent, ChatTitleUpdatedEvent) plus an OtherEvent fallback for forward compatibility — mirrors the Rust core's enum and the four event types the docs don't mention but the Rust PR captured from live traffic (workflow_started, ping, chat_finished, chat_title_updated)
  • message_id accepts either a JSON string or a raw number across all the payloads that carry it, matching the Rust SDK's defensive deserialization

• http.Client changes needed to support this:

  • Extracted buildRequest (shared request building/signing) out of Call
  • New CallSSE, returning the raw response body for streaming instead of buffering it
  • New WithRequestTimeout request option. AgentContext uses its own underlying *http.Client with no overall timeout (a full LLM turn can run long), and instead applies a dedicated 120s timeout to the blocking Conversation/Continue calls only — ConversationStream/ContinueStream are left unbounded (only ctx cancellation stops them), matching the Rust core's AGENT_REQUEST_TIMEOUT / unbounded-stream-reads design. Plain Workspaces/Agents GETs get a short (15s) timeout via the same option, since they're ordinary fast REST calls.

Test plan

• go build ./...
• go vet ./agent/... ./http/...
• agent/stream_test.go / agent/types_test.go — unit tests replaying the exact SSE fixtures and JSON payloads from the Rust PR's test suite, including the out-of-order chat_title_updated-after-workflow_finished case and the docs' succeeded/interrupted ConversationResponse examples
• http/client_test.go — httptest.Server-backed tests for CallSSE (success + error body) and WithRequestTimeout
• Live-tested against the real production API with a real account/workspace/agent: Workspaces → Agents → blocking Conversation → ConversationStream, confirming actual SSE event ordering (chat_started → workflow_started → ping → message chunks → chat_finished → ping → workflow_finished) matches what's implemented