Threaded dispatcher with backpressure support

[shikokuchuo]

Closes #454. Closes #581.

Summary

• Dispatcher moves from a separate R process to an in-process thread (Threaded dispatcher #581). The dispatcher logic, already reimplemented in C (nanonext 1.8.0), now runs as a thread within the host process. This eliminates inter-process communication overhead for dispatching and status queries, using inproc:// transport instead of IPC sockets. info() reads dispatcher state directly rather than sending a query-response round trip.

• New capacity argument for daemons() (Gating dispatcher based on uncollected results #454). Sets the maximum number of tasks (queued + executing) at the dispatcher. When the limit is reached, mirai() blocks until an existing task completes, providing backpressure to prevent unbounded memory growth. NULL (default) preserves existing unlimited queuing behavior.

Benefits

• Lower dispatch overhead by eliminating the separate process and IPC round trips
• Direct memory access for info() instead of send/recv query protocol
• Backpressure prevents unbounded memory growth when producers outpace consumers (ML dataloaders, DB write pipelines, Shiny promise chains)
• Simpler startup: no system2() call to spawn a dispatcher process, no multi-step synchronization handshake
• Removes the exported dispatcher() function and R/dispatcher.R — net deletion of ~80 lines

fyi @lionel- this is a significant milestone for mirai

[wlandau]

This is fantastic, @shikokuchuo! I notice a speedup in my long tests, and there info() no longer shows up in any of my profiling data. Really exciting work!

I do have a couple unit tests to update in crew because there are no longer any dispatcher processes, but that should be quick.

[shikokuchuo]

Revdeps coming in clean