Thread-safe python readers

[texodus]

This PR introduces a rewrite of perspective-python's concurrency model, enabling safe GIL-less+concurrent calls to Perspective's API, as well as additional assorted performance fixes (including a significant improvement to to_arrow() for both Python and Wasm). In addition to Perspective's internal parallelism, PerspectiveManager can now process requests concurrently by providing a thread pool dispatch to PerspectiveManager.set_loop_callback():

with concurrent.futures.ThreadPoolExecutor() as executor:
    manager.set_loop_callback(psp_loop.run_in_executor, executor)

• Fixes a regression in 2.2.1 (only this version) which disabled PSP_PARALLEL_FOR for all python builds.
• The semantics of async mode have changed. Before this PR, set_loop_callback() causes GIL unlock behavior, and calls to Perspective APIs from different threads will throw an exception as a protection against the engine's internal thread-unsafeness. With this PR, the GIL is released always, and Perspective's API implements its own internal locking. As a result, a PerspectiveManager can now be called without the GIL via multiple threads from Python, e.g. from a concurrent.futures.ThreadPoolExecutor, substantially increasing thread utilization on servers with many concurrent client requests.
• Apache Arrow reading/writing have been ported to utilize PSP_PARALLEL_FOR across columns.
• Removes numerous superfluous heap allocations from to_arrow(), improving serialization performance by ~50%:
• Fixes a bug in Python PerspectiveManager which caused some callbacks to become crossed when multiple clients are connected.

Historic to_arrow() wasm performance

Historic Python performance across a range of benchmarks (2.2.1 has a specific regression in PSP_PARALLEL_FOR):