fix(api): run native PDF extraction off the Node.js event loop

processPdf/detectPdf were synchronous napi calls. A large/complex PDF
(e.g. 181 pages, complex layout) blocked the calling pod's event loop
for its full extraction time (~25s observed in prod), freezing every
in-flight request on the process, failing liveness probes, and timing
out unrelated scrapes sharing the pod.

Both bindings now offload to tokio's blocking pool via spawn_blocking
(napi async fn), so extraction runs on a background thread and the
event loop stays responsive. Call sites just await the same functions.

Adds a regression test that fails on the old sync binding (measured
~1.25s of event-loop lag) and passes on the async one.