Modifications to bootstrapping to reduce memory usage

[ECharria]

Summary
Reduces memory usage of bootstrapping.py by using smaller integer dtypes for count and indicator matrices. Outputs are bit-identical to the previous implementation, counts and 0/1 indicators are exact integers, only their representation changes. Motivated by an OOM crash on a 100,517-spectrum run: the original code allocated several (N, N) float64 matrices (~75 GiB each) to hold integer counts.

Changes (in specreboot/bootstrapping/bootstrapping.py)

1. total_pair_counts, total_edge_support: float64 → uint16 (max value = B)
2. pair_counts, edge_support (per-iteration): int64 → uint8 (binary indicators)
3. mutual_topk result: float64 → uint8 (function only writes 0 or 1)
4. Same changes mirrored in _reconstruct_history
5. total_pair_similarities intentionally left as float64 to preserve precision of the similarity sum

Impact
At N = 100,517, each (N, N) matrix drops from 75.3 GiB (float64) to 18.8 GiB (uint16) or 9.4 GiB (uint8). The dataset that previously OOM'd at ~189 GiB RSS now runs to completion at ~111 GiB.
Assumes B ≤ 65,535 (uint16 capacity); comfortable for realistic SpecReBoot runs.

Verification

1. Equivalence test (B=20, N=200): mean_similarity and mean_edge_support bit-identical (max abs diff = 0.00e+00)
2. 100,517-spectrum GNPS bile acid dataset: completed successfully
3. 112,086-spectrum GNPS bile acid dataset: completed successfully