fix(cursor): normalize and smooth agent streams

[ssakone]

Summary

• emit Cursor thinking as OpenAI reasoning_content and keep reasoning separate in non-streaming responses
• normalize whitespace and control characters in tool call IDs so generated calls match returned tool results
• report non-stream token usage and avoid duplicate [DONE] markers in OpenAI streams
• coalesce bursty protobuf deltas at a 16 ms display cadence while preserving immediate first output and flushing before tool-call boundaries

Test plan

• gofmt -w internal/runtime/executor/cursor_executor.go internal/runtime/executor/cursor_executor_test.go
• go test ./...
• go test -race ./internal/runtime/executor -run 'Test(CursorStreamCoalescer|NormalizeToolCallID)' -count=10
• go build -o test-output ./cmd/server
• validate OpenAI and Anthropic streaming tool calls against an isolated server
• benchmark direct and downstream-proxied streams for cadence, first output, and completion latency