Nearest neighbors with trees perf decreased by debugging stats

Description

For ball_tree and kd_tree algorithms, some stats about the tree queries highly decrease the parallelization performances increase.

Those stats are:

• n_trims: queried points outside node radius
• n_leaves: leaves reached while querying
• n_splits: non-leaves queried nodes
• n_calls: num of computed distances

Those stats only seem useful for debugging, do not look like part of the official API (no documentation) and only 2 (personal) git repos use the method (get_tree_stats) to get them.

Deactivating them highly improves performances of associated algorithms.

Benchmark

Test of kneighbors function with default parameters and:

• samples dimension: 100
• fit: 10k samples
• kneighbors: 10k samples

(also tested openMP prange parallism but it does not improve perf)

=============
=== brute ===
=============
Joblib (loky) :
- n_jobs = 1 (MKL mono threaded) -> 2.6s
- n_jobs = 1 (MKL multi threaded, 40 threads) -> 1.9s
- n_jobs = 4  -> 4.0s
- n_jobs = 10 -> 3.5s
- n_jobs = 40 -> 3.5s

=================
=== ball_tree ===
=================
Joblib (loky) :
- n_jobs = 1  -> 10.9s
- n_jobs = 4  ->  7.7s
- n_jobs = 10 ->  6.8s
- n_jobs = 40 ->  3.8s

Joblib (loky) no stats:
- n_jobs = 1  -> 12.0s
- n_jobs = 4  ->  3.2s
- n_jobs = 10 ->  1.4s
- n_jobs = 40 ->  0.6s

OpenMP no stats:
- n_jobs = 4  ->  3.2s
- n_jobs = 10 ->  1.4s

===============
=== kd_tree ===
===============
Joblib (loky) :
- n_jobs = 1  -> 19.1s
- n_jobs = 4  ->  9.0s
- n_jobs = 10 ->  10.9s
- n_jobs = 40 ->  8.5s

Joblib (loky) no stats:
- n_jobs = 1  -> 19.0s
- n_jobs = 4  ->  5.1s
- n_jobs = 10 ->  2.2s
- n_jobs = 40 ->  1.0s

OpenMP no stats:
- n_jobs = 4  ->  5.1s
- n_jobs = 10 ->  2.2s