Wire GPU trace postprocessing into trace session

[jafioti]

Motivation

• Centralize Perfetto trace output tracking so GPU postprocessing can reuse the trace file created by the unified tracing initializer.
• Let GPU timing postprocessing accept path-like inputs and be invoked automatically when a Perfetto sink is configured.
• Provide a simple TraceSession API to flush/stop traces and expose the Perfetto output path to avoid duplicating Perfetto wiring in examples.

Description

• Add src/trace.rs with TraceOptions, TraceSink, TraceSession, init, and helper trace_file_path, and expose the module via pub mod trace;.
• Record the Perfetto trace file path in TraceSession via a new perfetto_path field and accessor perfetto_path() so other code can postprocess the same file.
• Change crates/luminal_cuda::record_exec_timings_to_file to accept impl AsRef<std::path::Path> and read/write using the provided path, and call it from examples/llama only when trace_session.perfetto_path() is Some.
• Emit structured tracing around search/egglog metrics and add per-op spans for native ops (src/graph.rs and src/op.rs) so traces include these events.

Testing

• Ran cargo fmt, which completed successfully.
• No additional automated builds or tests were executed.

---

Codex Task