fix: replace ProcessPoolExecutor with ThreadPoolExecutor (#552)

[parallelArchitect]

Fixes #552

Problem

Process-based parallelism duplicates zarr-backed data into each worker
via copy-on-write. On large volumes this causes severe memory inflation —
#552 reports 35–45 GB RSS for a 750 MB dataset with 4 workers.

Two additional hazards with the process-based approach:

1. tensorstore fork-safety: tensorstore runs internal C++ threads
   that are not fork-safe. A forked worker inherits locked mutexes from
   threads that don't exist in the child, causing deadlocks or segfaults
   (Segfault when using PyTorch Dataloader with multiple workers google/tensorstore#61). The spawn context avoided forking but added
   significant startup overhead per worker.

2. PyTorch thread limiting: torch.set_num_threads(1) and
   set_num_interop_threads(1) were required to prevent thread explosion
   across spawned processes. This artificially constrained PyTorch's
   internal parallelism.

Fix

Replace ProcessPoolExecutor with ThreadPoolExecutor. Threads share
memory — no copying, no fork-safety hazards, no spawn overhead.

numpy, scipy, and PyTorch all release the GIL during compute (FFTs,
Tikhonov solves), so threads achieve genuine parallelism on the
compute-heavy parts of the reconstruction pipeline.

Changes

• waveorder/cli/parsing.py: rename processes_option → threads_option,
  replace --num_processes with --num-threads. Deprecated aliases
  --num-processes / --num_processes retained for backward compatibility.
• waveorder/cli/apply_inverse_transfer_function.py: swap executor,
  remove spawn context and torch thread limiting.
• waveorder/cli/reconstruct.py: update to use threads_option and
  num_threads.
• waveorder/cli/utils.py: rename num_processes → num_threads in
  estimate_resources.
• waveorder/calib/calibration_workers.py: update keyword argument.

Testing

200 tests pass locally (cli_tests, api_tests, models).
waveorder/waveorder_simulator.py uses ProcessPoolExecutor for a
different purpose (CPU-bound Jones matrix computation, no zarr I/O) and
is intentionally left unchanged.

[ieivanov]

The PoolExecutor has been very carefully chosen to work with our biahub reconstruction pipeline. I advise not changing the default. We could offer an option to slot in a different executor if needed for certain applications.