Add --fake flag for e2e test session replay

[stanislavHamara]

Summary

• Add --fake <cassette-path> flag to replay recorded AI responses for E2E testing
• Extract VCR proxy logic into reusable pkg/fake package
• Refactor e2e/proxy_test.go to use shared code, removing ~110 lines of duplication

Test plan

• Build passes (task build)
• Lint passes (task lint)
• Manual test with curl (see below)
• E2E tests pass (task test)

Manual testing

Terminal 1 - Start server with fake mode:

./bin/cagent api \
  --fake ./e2e/testdata/cassettes/TestExec_OpenAI \
  --listen :8080 \
  --session-db /tmp/test-session.db \
  ./e2e/testdata/basic.yaml

**Terminal 2 - Test normal endpoints and AI replay:

# Create session and send message
SESSION_ID=$(curl -s -X POST http://localhost:8080/api/sessions \
  -H "Content-Type: application/json" -d '{}' | jq -r '.id')

curl -N "http://localhost:8080/api/sessions/${SESSION_ID}/agent/basic.yaml" \
  -H "Content-Type: application/json" \
  -d '[{"role": "user", "content": "What'\''s 2+2?"}]'

**Expected output (replayed from cassette, no real API call):

data: {"type":"agent_info","agent_name":"root","model":"openai/gpt-3.5-turbo","description":"A helpful AI assistant"}
data: {"type":"stream_started","session_id":"...","agent_name":"root"}
data: {"type":"agent_choice","content":"2","agent_name":"root"}
data: {"type":"agent_choice","content":" +","agent_name":"root"}
data: {"type":"agent_choice","content":" ","agent_name":"root"}
data: {"type":"agent_choice","content":"2","agent_name":"root"}
data: {"type":"agent_choice","content":" equals","agent_name":"root"}
data: {"type":"agent_choice","content":" ","agent_name":"root"}
data: {"type":"agent_choice","content":"4","agent_name":"root"}
data: {"type":"agent_choice","content":".","agent_name":"root"}
data: {"type":"stream_stopped","session_id":"...","agent_name":"root"}