PERF: Speed up permutation cluster tests ~15x via Numba JIT kernels

The permutation loop in spatio_temporal_cluster_1samp_test is the main
computational bottleneck for source-space cluster analyses. On fsaverage
ico-5 (~20K vertices, 15 timepoints, 307K tests), 2048 permutations
previously took ~14 seconds; this patch brings it to under 1 second on
a 16-core EPYC.

The key changes, roughly in order of impact:

- Fused Numba union-find (_st_fused_ccl) replaces the Python BFS in
  _get_clusters_st. Handles both spatial neighbors (CSR adjacency) and
  temporal self-connections in a single compiled pass, avoiding the
  overhead of the old _get_clusters_spatial + _reassign loop.

- Parallel permutation processing (_perm_batch_fast) fuses threshold
  scan + CCL + weighted bincount + argmax into one @jit(parallel=True)
  function with prange across permutations. Each perm gets its own
  pre-allocated work buffers to avoid data races.

- Batched fused t-test (_batched_fused_ttest) reads X_T once from DRAM
  and computes t-statistics for 32 permutations, amortizing memory
  traffic ~32x. Uses prange across variables.

- Single-perm fused t-test (_fused_ttest) replaces the numpy sequence
  (dot product + 8 elementwise ops) with one prange loop.

- Compact-graph CCL in _get_components: builds a subgraph of only the
  supra-threshold vertices before calling connected_components, so CCL
  operates on ~1K vertices instead of ~20K.

- Vectorized _pval_from_histogram with np.searchsorted (O(n log n)
  instead of O(n * n_perms)).

- Vectorized _get_1samp_orders via bit-shifting instead of per-element
  np.fromiter(np.binary_repr(...)).

- Pre-computed CSR arrays from _setup_adjacency, threaded through to
  _find_clusters and the permutation loop to avoid redundant
  neighbor-list-to-CSR conversion.

All fast paths are gated behind has_numba checks and identity checks on
the stat function (only activates for the built-in ttest_1samp_no_p and
f_oneway). Non-Numba fallback paths are unchanged.

Tested on EPYC 9R14 (16 vCPU), fsaverage ico-5, 15 subjects, 2048
permutations: 0.37 ms/perm wall time vs ~5.7 ms/perm baseline.

Author: sharifhsn

doc/changes/dev/XXXXX.newfeature.rst

@@ -0,0 +1 @@
+Speed up :func:`mne.stats.spatio_temporal_cluster_1samp_test` permutation loop by ~15x via Numba JIT kernels for the t-test, threshold extraction, and connected-component labeling steps, by :newcontrib:`Sharif Haason`.

doc/changes/names.inc

@@ -307,6 +307,7 @@
 .. _Sena Er: https://github.com/sena-neuro
 .. _Senwen Deng: https://snwn.de
 .. _Seyed Yahya Shirazi: https://neuromechanist.github.io
+.. _Sharif Haason: https://github.com/sharifhsn
 .. _Sheraz Khan: https://github.com/SherazKhan
 .. _Shresth Keshari: https://github.com/shresth-keshari
 .. _Shristi Baral: https://github.com/shristibaral