feat: add SOAP-based symmetry-equivalent site detection

[sandipde]

Add a second method for identifying symmetry non-equivalent adsorption sites using SOAP (Smooth Overlap of Atomic Positions) descriptors from the dscribe package with hierarchical clustering, as an alternative to ASE's SymmetryEquivalenceCheck.

Core implementation (autoadsorbate/utils.py):

• _compute_soap_site_vectors(): compute SOAP descriptors for candidate sites
• _soap_similarity_matrix(): pairwise cosine similarity matrix
• _filter_unique_sites_by_soap(): cluster sites via hierarchical clustering or greedy merging, with configurable similarity_threshold (default 0.99)

Unified API (autoadsorbate/autoadsorbate.py):

• Surface.get_nonequivalent_sites() accepts method="ase"|"soap"
• Surface.sym_reduce() accepts method="ase"|"soap"
• New methods: get_soap_similarity_matrix(), get_nonequivalent_sites_soap(), soap_reduce()

Tests (tests/test_all.py):

• 13 tests covering SOAP similarity matrix, hcluster/greedy modes, soap_reduce, method switching, ASE vs SOAP comparison on Ni/Ru slab, timing benchmarks

Documentation (README.md):

• SOAP method usage section with API examples
• Threshold tuning guide
• Dedicated ASE vs SOAP benchmark section with:
  • Method comparison table (ASE: 16 sites/8.4s vs SOAP: 8 sites/0.016s)
  • Detailed breakdown by site type (8 SOAP clusters)
  • Site comparison plot (site_comparison.png)
  • SOAP distance histogram (soap_histogram.png)
  • Discussion of why SOAP outperforms ASE on broken-symmetry surfaces

Scripts:

• scripts/compare_sites.py: detailed ASE vs SOAP comparison
• scripts/plot_sites.py: site visualization with color-coded SOAP clusters
• scripts/site_comparison.png: generated comparison image
• scripts/soap_histogram.png: SOAP distance distribution plot

On Ni(111) 3x3 slab with one Ru dopant:

• ASE finds 16 unique sites in ~8.4 s
• SOAP finds 8 unique sites in ~0.016 s (~525x speedup)