Async perf paths do not wake up for in-process or mock-only data sources

[utpilla]

Follow up to #279 (comment)

PR #279 added PerfSession::perf_fds() so consumers can drive parsing from kernel fd readiness (for example epoll(2) or tokio::io::unix::AsyncFd).

For sources that do not own per-CPU perf fds (MockData and InProcessRingBuf), perf_fds() returns an empty list. A consumer wired purely to fd readiness will therefore never wake up on such sources.

The current contract is that the consumer picks the source and knows whether the fd-driven async pattern applies. Sources without fds are expected to be driven by PerfSession::parse_all, parse_for_duration, or parse_until, as they are today.

Possible future directions if a real use case appears:

• Give InProcessRingBuf a std::task::Waker slot (for example Arc<Mutex<Option<Waker>>> shared with the writer). The reader stashes its waker on Poll::Pending, and InProcessRingBufWriter::write calls wake() after publishing. This is std-only, works on any async runtime, and avoids a kernel-side wakeup object.
• Alternatively, add an eventfd (or similar) so perf_fds() can return a real fd for in-process sources.
• Add a timeout boundary to the async drain paths so they fall back to a periodic drain when no fds are present.

MockData is a test fixture, so the sync drain methods (parse_all, etc.) are likely the right answer there regardless.