Implement embed v2 data-loading pipeline + H100 benchmark harness

[clemsgrs]

Summary

This PR implements the new embed v2 data-loading pipeline focused on single-node H100 throughput while keeping v1 compatibility.

Core changes

• Add speed.embedding_pipeline (v1/v2) in embedding config.
• Add adaptive sharding for v2:
  • rank_sharding_mode=auto -> slide sharding when pending_slides >= world_size, else tile fallback.
• Add loader tuning knobs and heuristics:
  • num_workers_embedding supports "auto"
  • prefetch_factor_embedding, persistent_workers_embedding, pin_memory_embedding, loader_batch_timeout_sec, storage_mode
• Refactor tile loading to cache one worker-local WSI handle in TileDataset.
• Remove embed warmup dry-run batch and switch to lazy writer initialization from first real batch.
• Add optional perf timing logs per slide (speed.log_perf_embedding).
• Keep output compatibility (features/<slide>.pt by default).

Benchmark harness

• Add scripts/benchmark_embed_v1_v2.py to compare embed-only throughput for v1/v2 on GPU sets like 1,4,8.
• Script expects an existing tiling output with process_list.csv and coordinates/.

New/updated tests

• test/test_dataset_worker_cache.py
• test/test_rank_sharding_lpt.py
• test/test_embed_pipeline_mode.py
• test/test_regression_bugfixes.py (new check for v2 config key usage)
• CI regression job now runs all test_*.py files.

Validation

• python3 -m unittest discover -s test -p 'test_*.py'
• python3 -m compileall slide2vec/embed.py slide2vec/data/dataset.py scripts/benchmark_embed_v1_v2.py
• python3 scripts/benchmark_embed_v1_v2.py --help

Notes

• slide2vec/hs2p submodule code was not modified in this PR.