Add return_sig and decouple significance compute from display

[lshpaner]

Summary

Extends flex_corr_matrix with p-value return support and fixes two significance-related behavior bugs. All changes are backward compatible at default arguments.

Changes

• Add return_sig to return the pairwise p-value matrix as a DataFrame, mirroring return_corr. When both return_corr and return_sig are set, returns a (corr_matrix, pval_matrix) tuple. Return annotation loosened to Optional[Union[pd.DataFrame, Tuple[pd.DataFrame, pd.DataFrame]]].
• Fix _stars significance gate. The star gate was hardcoded at p >= 0.05, so significance_level had no effect in "stars" mode despite the docstring claiming otherwise. Now gates on significance_level, so stars and mask share one threshold. Identical output at the 0.05 default.
• Decouple p-value computation from display. Introduced a _need_pvalues trigger (show_significance or filter_significance is not None or return_sig) so p-values are computed whenever any consumer needs them, while the stars/mask overlay stays gated on show_significance.
• Remove silent show_significance auto-enable. filter_significance no longer overrides an explicit show_significance=False. It now filters variables without forcing the overlay, so filter_significance=0.05, show_significance=False produces a plain filtered heatmap.
• Plot-suppression gate now covers both return flags: return_corr or return_sig suppresses the plot by default; show_plot=True re-enables.
• Docstring reconciled: new return_sig entry, corrected significance_level / filter_significance / show_plot descriptions, updated Returns for the three-way return.

Add p_adjust, return_n, show_n to flex_corr_matrix

• p_adjust: bonferroni / fdr_bh, corrected over unique pairs (m = k(k-1)/2), applied before stars, mask, filter_significance and return_sig consume p-values
• return_n: pairwise sample-size matrix; n_format; toggles counts vs percent of frame
• show_n: plots the n matrix as a coverage heatmap; mutually exclusive with show_significance
• cmap/vmin/vmax/cbar_label now resolve by mode
• Fix: filter_significance dropped the -0.00 zeroing
• 40 tests, all mutation-checked; restores the pulled _stars significance_level test via gcf().axes

Tests

Added coverage for return_sig (single return, tuple return, plot suppression, diagonal, filter/threshold reflection). Note: the _stars-gate regression test was pulled pending a fix to axis-annotation readback, so the gate behavior is currently unpinned.

_________ coverage: platform linux, python 3.12.7-final-0 _________

Name                                     Stmts   Miss  Cover
------------------------------------------------------------
src/eda_toolkit/__init__.py                 19      2    89%
src/eda_toolkit/_data_manager_utils.py      28      0   100%
src/eda_toolkit/_plot_utils.py             185     23    88%
src/eda_toolkit/art.py                      54      3    94%
src/eda_toolkit/data_manager.py            846     84    90%
src/eda_toolkit/plots.py                  1314    236    82%
------------------------------------------------------------
TOTAL                                     2446    348    86%