fix: pin BLAS to one thread in PWM kernels to stop thread explosion

[aviezerl]

Problem

extract_pwm() (and compute_pwm() / compute_local_pwm() / calc_seq_pwm()) open thousands of OS threads and can fail outright on core-limited machines. Reported via iceqream::infer_trajectory_motifs(), whose default path runs extract_pwm().

Root cause is nested parallelism: these kernels run a small per-sequence BLAS dgemm (PWMWorker::compute_matrix_scores) inside an RcppParallel/TBB parallelFor over sequences. With a multi-threaded BLAS (MKL or OpenBLAS), each TBB worker spawns its own BLAS thread team, so one call opens n_threads × n_blas_threads threads:

setThreadOptions	OS threads (128-core node, MKL)
16	1025 (64×16+1)
128	8193

At the set_parallel() default (~102 threads) that's ~6500 threads for a single extract_pwm. It oversubscribes badly and, on a cluster job with a thread/process (cgroup pids) limit, fails with a thread-creation error.

(A bare RcppParallel parallelFor opens exactly n_threads; the multiplier is entirely the inner threaded dgemm. OMP_NUM_THREADS=1 at launch also collapses it - confirming the diagnosis.)

Fix

New helper local_serial_blas() pins the BLAS to one thread for the duration of these kernels and restores the previous count on exit. The per-sequence dgemm is tiny and the parallelFor over sequences is the right parallelism granularity, so the inner BLAS should never be threaded here.

compute_pwm / compute_local_pwm / calc_seq_pwm:
    local_serial_blas()   # BLAS -> 1 thread, restored on return
    ...parallelFor...

Verification (MKL)

• 16-thread extract_pwm: 1025 → 17 OS threads; compute_pwm / compute_local_pwm likewise.
• Results identical (all.equal TRUE); throughput same or slightly better (the threaded dgemm was pure overhead at this size).
• BLAS thread count restored after the call, so the rest of the session keeps threaded BLAS.
• Full testthat suite passes.

Adds RhpcBLASctl to Imports (the portable cross-BLAS thread control - RhpcBLASctl::blas_set_num_threads). Bumps to 0.0.10.

https://claude.ai/code/session_01KfEW7AGsSzjhMqiY8htdnq