speedup: vectorise Monte Carlo permutation test in fisher_test_cci

[Copilot]

The Monte Carlo permutation loop inside fisher_test_cci called fisher_exact 1 000 times per cell pair via a plain Python for loop, with np.random.seed(42) reset on every iteration (making all permutations across rows identical). perm_p_value was computed but never stored.

Changes

• Bug fix – RNG seed reset in loop: replaced np.random.seed(42) inside the inner loop with a single np.random.default_rng(42) created once per function call.
• Vectorised permutation generation: all B permutations are now produced in one NumPy call instead of B sequential calls:

  permuted = rng.permuted(np.tile(flat, (B, 1)), axis=1).reshape(B, 2, 2)

• Deduplicated fisher_exact calls: 4 values yield at most 4! = 24 unique 2×2 arrangements; np.unique(..., return_inverse=True) reduces ~1 000 calls per row to ≤ 24.
• Vectorised p-value counting: replaced the manual count += 1 loop with (perm_pvals <= p_value + _PVAL_TOL).sum().
• perm_p_value now stored: added as a perm_p_value column in the output stats DataFrame.
• Eliminated code duplication: extracted a _pairwise_stats helper shared by both branches of fisher_test_cci.