Guard numba's thread pool against concurrent entry from thread pools

[cmdupuis3]

Closes #1798

Overview

A parallel=True kernel called from a thread pool -- dask's threaded scheduler, typically -- nests numba's pool under it. What that does depends on numba's threading layer:

• tbb composes; nothing to do.
• omp starts a full team per calling thread: 179 OS threads from a 12-thread pool.
• workqueue aborts the process on concurrent entry. That is live today: a chunked neighborhood reduction (the target="parallel" gufunc runs inside apply_ufunc(dask="parallelized")) kills the interpreter. It predates the nogil change; it aborts with the GIL held too.

numba_pool() in uxarray/utils/parallel.py guards each call by layer: a lock under workqueue (on every thread, since a main-thread call racing a dask task aborts too), numba.set_num_threads(1) off the main thread under omp, restored afterwards because the setting is thread-local. The 14 parallel=True kernels move to @parallel_njit, which is njit plus that guard; the neighborhood gufunc wraps its call in it.

The two remedies are per layer on purpose. Capping does not stop the workqueue abort, and capping under the lock would run every pool-thread call one at a time on one thread.

Chunked neighborhood percentile, 64 steps, best of 3:

layer        before            after
tbb          139ms             139ms
omp          179 OS threads    47 OS threads
workqueue    abort             ~1.1s

The workqueue time is the layer, not the lock: main takes 1176ms under dask's sync scheduler.

Test suite: 983 passed, 1 skipped. test_plot_with_features fails identically before and after (matplotlib figure size, unrelated).

PR Checklist

General

• An issue is created and linked
• Added appropriate labels (if your uxarray repo permissions allow it)
• Filled out Overview and Expected Usage (if applicable) sections

Testing & Benchmarking

• There is adequate test coverage of changes from this PR (add new tests if needed)
• If this PR could affect performance, ran ASV benchmarks and confirmed they show expected behavior (add a new benchmark if necessary)

Documentation and Examples

• Docstrings updated with any function changes, and included in all new functions
• User (public) functions added to docs/api.rst; internal (private) function names start with an underscore (_)

AI Disclosure

AI Usage: Claude Opus 5.5

• I have tested and take responsibility for all AI-generated content in my PR.