Run the threaded pencil chunks with Polyester's @batch

[pankgeorg]

One @batch chunk per worker instead of one Threads.@spawn task per chunk in _foreach_pencil (src/plan.jl); output bit-identical, new dependency Polyester.
compare3 vs integration/all @ 901692b, aarch64 Neoverse-N1, 117 N-d/batched cases per thread count: 1 thread 0.99×, 4 threads 1.05×, 16 threads 1.16× geomean (0.95–1.91), no case slower by >5 % — e.g. batched 4096×64 400→216 µs, 1024×64 81→43 µs, 64³ 1.10→0.83 ms, 128³ 11.7→7.9 ms; DRAM-bound 64×65536 at parity. x86-64 (Core Ultra 7 165H, 6P+8E, benchmark/x86-64/compare3/polyester/ on bench/x86-compare3): 32K–256K-element band 1.32× at 4 threads, 1.79× at 8, 2.32× at 12 — @spawn saturates at 8 threads (4.5×), @batch keeps scaling (6.9× at 12); <32K and DRAM-bound bands at parity, checksums identical. Tables: benchmark/compare3/aarch64/polyester/COMPARE3.md on bench/compare3.
Tested: full suite green on 4 threads.