ENH perf: numeric optimisation in nan_euclidean_distances, GaussianMixture and KNeighborsRegressor.predict - #34277

Three independent speedups found in a perf review:

1. nan_euclidean_distances: compute the present-feature-count matrix with
   a float matmul instead of an integer one. NumPy has no BLAS path for
   integer matmul, so the boolean masks are cast to the working dtype
   first. Counts are small integers, exactly representable, so the result
   is unchanged. Speeds up KNNImputer and pairwise_distances/NearestNeighbors
   with metric="nan_euclidean".

2. GaussianMixture (tied covariance) _estimate_log_gaussian_prob: replace
   the per-component loop with a single BLAS expansion of the squared
   distance ||Xp - mu_proj||**2, the same form the diag and spherical
   branches already use. Not bit-identical to the loop, but a full tied
   fit is unchanged in practice (identical means_, lower_bound_ within ~1e-15).

3. KNeighborsRegressor.predict: fold the per-output distance-weighting
   loop into a single einsum that gathers the neighbor targets once.
   Output identical; speeds up many-output multi-target regression,
   neutral for single-output.

Benchmarks (median, baseline vs branch, real APIs):
  nan_euclidean_distances (n=10000, ny=2000): d=50 3.4x, d=200 7.8x, d=500 10.7x
  KNNImputer.fit_transform (15% missing):     1.20-1.25x
  GaussianMixture(tied).fit:                  1.9x (c=5) to 5.5x (c=40 / d=100)
  KNeighborsRegressor.predict:                1.0x (1 output) to 2.1x (1000 outputs)