MAR & time-gap imputation improvements, 4.5x faster imputation pipeline

[CyrilJl]

Summary

This PR improves TimeSeriesImputer for missing-at-random values and contiguous timestamp gaps, then makes the whole imputation pipeline substantially faster and lighter in memory — with imputation results unchanged (verified elementwise, differences at float32 rounding level only).

Imputation improvements (cf66f9e)

• TimeSeriesImputer can now infer a regular DatetimeIndex frequency, reinsert missing timestamp rows inside the observed range, and impute them.
• Optional deterministic calendar/trend features (add_time_features=True) that stay fully observed through contiguous gaps, so blocks of missing timestamps can be filled even when all lag features are NaN.
• Training subsets that end up identical across missingness patterns are fitted once and reused.

Performance (b12ff96, df2e12c)

• Gram-based ridge fast path: with the default FastRidge regressor, per-pattern models are solved from an augmented [X, y, 1] Gram matrix accumulated once per column over the globally complete rows plus a small per-pattern correction, instead of materializing and refitting a large row subset for every missingness pattern (~500 patterns/column on the benchmark). Custom regressors and categorical targets keep the exact estimator.fit path.
• Cheaper pattern handling: per-pattern complete training rows are found from a CSC layout of NaN positions with epoch-stamped scratch buffers — cost scales with the NaN count in excluded columns instead of n_rows × n_usable_cols.
• Fixed-cost reductions: fused masked normalization stats (replacing nanmean/nanstd full-matrix copies, applied in place), feature-selection scoring rewritten from masked sums (same math, no pre-impute copy), lag/lead features built in one preallocated numpy matrix instead of df.shift + pd.concat, and output sliced before the DataFrame wrap so pandas CoW does not copy the full feature matrix.
• Memory fix: nan_positions over-allocated two m·n uint32 buffers kept alive by the returned slices; NaN indexing now allocates exactly what it needs. Dead helpers and an unused NaN-position pipeline (computed per call, never read) were removed.

Before / after

Reference benchmark: perf/timeseries_imputer_benchmark.py — 30,000 rows × 250 series (15-min frequency), 2% MAR + 8 series with 3,000-row outage blocks, 3 target series, n_nearest_features=35, lags=(1, 2, 3, -1, -2, -3). "Before" is this branch prior to the optimization commits; same machine, same environment.

Metric	Before	After	Change
Wall time (steady-state run)	5.88 s	1.32 s	4.5× faster
Time per imputed column	~0.87 s	~0.16 s	5.4× faster
Peak RAM (peak working set)	1,746 MB	1,181 MB	−32%
RSS retained after the call	512 MB	317 MB	−38%
Imputation MAE on masked cells	0.31700122	0.31700116	unchanged

The all-columns DataFrame path benefits as well (scripts/multivariate_benchmark.py, same environment): Diabetes MAR 10% drops from 1.20 s to 0.39 s with identical quality metrics on every dataset/pattern combination.

Quality verification

• Elementwise comparison of imputed values before/after on three scenarios (time series MAR + blocks, multivariate full-features, multivariate with n_nearest_features + min_samples_train): max relative difference ~1e-4, MAE identical to 7 decimals.
• scripts/multivariate_benchmark.py run on baseline and optimized code in the same environment produces identical metrics (RMSE/R²/accuracy) on all datasets.
• New regression tests pin the fast paths to reference implementations: complete_rows_excluding vs a plain-numpy oracle, the Gram path vs a materialized FastRidge, and the rewritten scoring vs the pre-impute formulation.

Testing

• pytest — 39 passed.
• ruff check / ruff format --check — clean.
• Benchmarks and profiles: perf/timeseries_imputer_benchmark.py, perf/timeseries_imputer_profile.py, scripts/multivariate_benchmark.py.