fix(cache): hash numpy scalars in cache keys

[Yigtwxx]

Problem

sha256_hexdigest falls back to json.dumps for anything that isnt bytes/str/ndarray/list/dict. numpy scalars (np.int64, np.float32, np.bool_) arent json serializable so the cache key raises. easy to hit with a param sweep:

for n in np.arange(5, 50, 5):
    compute_projection(df, inputs="v", umap_args={"n_neighbors": n})
# TypeError: Object of type int64 is not JSON serializable

np.float64 works only because it subclasses float.

Fix

convert np.generic with .item() before hashing. np.int64(15) and 15 now give the same key. keys that already worked (incl np.float64) dont change so existing caches stay valid. added a test in test_cache.py.