Phase 3: Tier P parallel execution via girder actors

[zacharywhitley]

Background

Design: docs/native-execution-and-parallelism.md §5, §6.
Depends on #1.

GIL-free parallelism by running many isolated CPython-WASM instances as girder
actors (each its own interpreter/GIL/memory), coordinated by message passing.

Tasks

• CPython-WASM → girder adapter: export girder's turn-actor world
  (init/handle) and dispatch handle(msg) to a Python entrypoint
  (the offload shape adapted to the actor ABI).
• Reproduce the parallel_mapreduce pattern (see
  ~/git/girder/crates/girder-runtime/examples/parallel_mapreduce.rs) with a
  CPython payload; spawn K actors, fan tasks, collect results.
• Add the arrow codec for array/frame payloads.
• Use girder SIR (Shared Immutable Regions) to broadcast large read-only
  inputs zero-copy.
• Measure multicore scaling and the data-movement cost.

Acceptance

A CPU-bound Python map runs across N girder actors with measured speedup vs a
single instance, and a large read-only input shared via SIR.

Notes / deps

• Sibling repo ~/git/girder (native host runtime; not browser — see design §7).
• Requires the CPython→turn-actor adapter above as a prerequisite.

[zacharywhitley]

Progress: actor adapter + parallel worker pool landed (03e31b6)

The verifiable Python-side core of Tier P is in place and green in CI (reference-worker, 19 tests):

• actor.py — girder turn-actor adapter (init / handle(message) -> bytes),
  the entry the eventual CPython-WASM girder actor exports; bridges to
  worker.run via the shared frames.
• pool.py — WorkerPool.map(env, tasks) fans tasks across N resident workers
  and collects results in order. Each worker is a separate process (own
  interpreter, own GIL, own memory) — exactly girder's shared-nothing model, where
  the GIL never serializes work across workers.
• test_pool.py — actor handle round-trip + a parallel map proven to run
  across distinct processes (distinct pids). The CPU-speedup measurement is the
  deferred benchmark (needs multiple cores + the girder runtime), not a unit test.

Remaining (needs the girder runtime)

• Componentize CPython-WASM to export girder's turn-actor world (init/handle).
• Run WorkerPool against real girder actors (swap the client_factory).
• Add the arrow codec; use girder SIR for large read-only inputs (zero-copy fan-out).
• Benchmark multicore scaling (the GIL-free speedup).