Add CodSpeed performance benchmarks and CI workflow

[codspeed-hq]

What this adds

Continuous performance measurement for Feature-engine with CodSpeed, so performance regressions are caught on pull requests instead of after a release.

Benchmark suite (benchmarks/)

109 benchmarks written with pytest-codspeed, one module per transformer family. fit and transform are benchmarked separately because their performance profiles are very different (learning parameters vs. applying them):

File	Covers
test_imputation.py	Mean/median, arbitrary, end tail, categorical and random sample imputers, missing indicator, drop missing data
test_encoding.py	Count, ordinal, mean (with and without smoothing), WoE, one-hot, rare label, decision tree and string similarity encoders
test_discretisation.py	Equal frequency, equal width, geometric width, arbitrary and decision tree discretisers
test_outliers.py	Winsoriser (gaussian, iqr, quantiles, mad), arbitrary capper, outlier trimmer
test_transformation.py	Log, power, reciprocal, Box-Cox, Yeo-Johnson, arcsin, arcsinh, mean normalisation
test_creation.py	Math, relative, cyclical and decision tree features
test_datetime.py	Datetime feature extraction (including parsing dates stored as strings), subtraction, ordinal
test_timeseries.py	Lag, window and expanding window features
test_selection.py	Constant, duplicate, correlated, high PSI, information value, target encoding and single feature performance selectors
test_variable_handling.py	The find_*/check_* helpers, which run inside every transformer's fit
test_pipeline.py	Two end-to-end pipelines plus MatchVariables and MatchCategories

Details worth noting:

• Data is synthetic and built in session-scoped conftest.py fixtures (numerical, strictly positive, bounded, categorical with rare labels, and datetime variables, with a variant containing missing values). Data generation therefore never lands inside a measured section.
• Three dataframe sizes are used so every benchmark stays in the millisecond range under CPU simulation: 10k rows for the vectorised transformers, 1k rows for the ones that train models under the hood, 500 rows for the row-wise ones (MathFeatures).
• Only non-deprecated public classes are used, so the suite produces no FutureWarning and will not break at the 2.1.0 removals.

CI workflow (.github/workflows/codspeed.yml)

Runs on pushes to main, on pull requests, and on workflow_dispatch (needed by CodSpeed to backfill initial data). It uses CodSpeedHQ/action@v5 in simulation mode on ubuntu-latest, with OIDC authentication (id-token: write), so no token secret is required.

Simulation mode was chosen because the benchmarks are CPU-bound: it gives deterministic, hardware-agnostic measurements with under 1% variance and produces flame graphs for each benchmark.

Other changes

• pyproject.toml: added testpaths = ["tests"] to the pytest configuration. Without it, a bare pytest (as used by the codecov tox environment) would also collect benchmarks/, which needs pytest-codspeed installed. Running pytest benchmarks/ explicitly still works.
• README.md: added the CodSpeed badge to the Testing row.

Verification

The suite was run locally in the same mode as CI:

codspeed run --mode simulation -- pytest benchmarks/ --codspeed
109 passed in 816.36s

All 109 benchmarks pass, and a plain pytest benchmarks/ --codspeed walltime run completes in about 7 minutes. Expect the CI job to take roughly 15 minutes, most of it CPU simulation overhead. flake8, black and isort were run on the new files.

Next steps

• Merge this PR so the main branch gets a performance baseline. Until then, PR reports have nothing to compare against.
• Once a baseline exists, CodSpeed comments on each PR with the benchmarks that moved, and flame graphs are available per benchmark to see where the time goes.
• Extend the suite as needed: add a benchmark next to the relevant module whenever a transformer's performance matters, following the guidance in benchmarks/README.md.

[codspeed-hq]

Congrats! CodSpeed is installed 🎉

🆕 109 new benchmarks were detected.

You will start to see performance impacts in the reports once the benchmarks are run from your default branch.

Detected benchmarks

• test_cyclical_features_transform: 76.3 ms
• test_decision_tree_features_fit: 1.3 s
• test_decision_tree_features_transform: 63.2 ms
• test_math_features_transform[func0]: 1.6 s
• test_math_features_transform[func1]: 2.3 s
• test_relative_features_transform: 40.4 ms
• test_datetime_features_from_string_transform: 56.3 ms
• test_datetime_features_transform[all]: 350.4 ms
• test_datetime_features_transform[basic]: 195.2 ms
• test_datetime_features_transform[default]: 227.7 ms
• test_datetime_ordinal_transform: 983.6 ms
• test_datetime_subtraction_transform: 180 ms
• test_arbitrary_discretiser_transform: 60.2 ms
• test_decision_tree_discretiser_fit: 1.8 s
• test_decision_tree_discretiser_transform: 65.4 ms
• test_equal_frequency_discretiser_fit: 151.2 ms
• test_equal_frequency_discretiser_transform[False]: 54.5 ms
• test_equal_frequency_discretiser_transform[True]: 2.1 s
• test_equal_width_discretiser_fit: 110.5 ms
• test_equal_width_discretiser_transform: 54.5 ms
• ...

ℹ️ Only the first 20 benchmarks are displayed. Go to the app to view all benchmarks.

---

[solegalli]

closes #964 #592 #622