BUG: Fix divide-by-zero and over-generation in KMeansSMOTE

[saitejabandaru-in]

Summary

Fixes #1186

KMeansSMOTE has two numerical issues that cause it to crash or generate the wrong number of samples:

1. When computing cluster sparsity (_find_cluster_sparsity), it divides by non_diag_elements. If X contains only one sample (or identically 1 valid point), non_diag_elements is 0, causing a divide-by-zero error.
2. When determining cluster_n_samples, it uses math.ceil(). If there are many valid clusters, taking the ceiling of every fractional sample count guarantees that the sum of generated samples will significantly exceed the requested n_samples.

Fix

1. Add a check if non_diag_elements == 0: return 0.0 in _find_cluster_sparsity.
2. Change math.ceil to round for cluster_n_samples to minimize overshoot. (Any exact sample distribution can be further refined, but round is standard for integer weighting).