add Gram matrix fast path for pairwise_distance

[kevinkorfmann]

Summary

• Add Gram matrix fast path for pairwise_distance() with euclidean/sqeuclidean metrics
• Uses d²(i,j) = ||x_i||² + ||x_j||² - 2*x_i·x_j via a single X @ X.T (n x n) instead of materializing all
  (n_pairs, n_variants) difference vectors
• Falls back to batched pairwise approach for cityblock metric or when missing data is present
• ~13x speedup, exact correctness (0.0 error vs scipy reference)

Benchmarks (NVIDIA A100, 200 haplotypes x 50k variants)

n_haps	n_snps	Before (ms)	After (ms)	Speedup
200	50,000	~75	5.8	~13x
200	100,000	~150	17.1	~9x

Test plan

• pixi run pytest tests/ -k 'decomposition or pca or pcoa or distance' (26 passed)
• pixi run python debug/bench_pairwise_distance.py

[kevinkorfmann]

Rebased onto the latest main and simplified the implementation. Euclidean and squared-Euclidean distances now use the memory-safe chunked Gram path, while missing-data and cityblock cases retain the existing fallback.

[nspope]

Sorry, this was rewritten to be multiallelic safe in #142 which is on the big multiallelic fix branch that's going to get merged in soon by #184. Can you wait until this is in, as it'll likely have merge conflicts?