…reader with hs2p

- restore supertile-reordered tile embeddings to coordinate order after
  forward pass (single-GPU and multi-GPU paths both go through
  _compute_tile_embeddings_for_slide which now applies argsort inverse)
- cap num_workers and num_preprocessing_workers to SLURM-allocated CPUs
  in ExecutionOptions.__post_init__ (checks SLURM_CPUS_PER_TASK,
  SLURM_CPUS_ON_NODE, SLURM_JOB_CPUS_PER_NODE); same env vars added to
  _slurm_cpu_limit() in inference.py
- drop legacy cfg.speed.num_workers fallback in ExecutionOptions.from_config;
  fix log message to say num_dataloader_workers instead of num_workers
- add suppress_c_stderr context manager in log_utils and wrap
  wholeslidedata imports to silence ASAP spdlog noise at import time
- bump hs2p minimum to 2.5.0 in all three locations (setup.cfg x2,
  requirements.txt); move PyTurboJPEG to base deps
- add tile-level embedding ground-truth fixture and assertion in
  test_output_consistency; add scripts/generate_gt.py to regenerate fixtures

Co-Authored-By: Claude Sonnet 4.6 <[email protected]>