In place shuffle dedup

Here is an AI generated text. Numba's Generator is very slow it turns out and we can write our own. But I will first wait for #1023 to be merged. It is bit-reproducable so it's not a breaking change. Should probably also open an issue about this in numba repo

Numba's Generator.shuffle overload is about twice as slow as the primitive it is built on: it copies through a 0-d array buffer for every swap (numba/np/random/generator_methods.py:188-196, buf[...] = z[j, ...] plus two more slice assignments) where a scalar swap would do, which costs roughly as much as the random draws themselves. Replacing it with an inline Fisher–Yates loop over the same random_interval draws, in the same descending order, is bit-identical — same permutation, same number of draws consumed, so a shared generator stays in step — and measures 2.0x faster at n=10,000 (0.080 ms vs 0.161 ms) and 2.2x at n=100,000 (0.727 ms vs 1.566 ms); it also beats NumPy's own Generator.shuffle, so this is worth an upstream numba issue independently. Squidpy has exactly two in-numba shuffle sites, _ligrec._score_permutations and _nhood._shuffled_labels, both shuffling a label array inside a prange loop from a typed list of generators, so both get the win; in nhood_enrichment the shuffle is roughly 70% of per-permutation work, and the indexed variant additionally removes the gather/scatter through group_indices for the library_key path (2.3-2.7x there) and makes the single-group special case redundant, so it can be deleted along with its test. The one thing to declare in the PR: it imports random_interval from numba.np.random.random_methods, a numba internal — if numba moves it the import fails loudly at module load, and the bit-identity tests would catch any change in behaviour.