option to re-use cv splits during tuning

[janrth]

At the moment the CV splits are created newly in each tuning trial, which can be slow for large datasets.
This PR adds an option to reuse CV splits across tuning trials.

When enabled, the splits are computed once in auto.py and passed into the optimization objective, where they are reused for all tuning trials. The default behavior is unchanged: if the option is disabled, CV splits are still generated inside each trial as before.

The original behavior is intentionally kept because this introduces a RAM vs CPU trade-off. Reusing splits can significantly reduce runtime when running many trials, but it keeps all train/validation splits in memory for the duration of the tuning run, which may increase peak memory usage on very large datasets.

To ensure correctness, a test was added using a deterministic model, verifying that predictions are identical when reusing CV splits versus recomputing them each trial.

When tested on the example below I saw a speed-up of 1.14x when re-using CV compared to the current implementation:

from utilsforecast.data import generate_series

series = generate_series(n_series=20000, freq='D', min_length=722, max_length=725, equal_ends=True)


# Setting up model
def my_lgb_config(trial: optuna.Trial):
    return {
        'learning_rate': 0.05,
        'verbosity': -1,
        'num_leaves': trial.suggest_int('num_leaves', 127, 128, log=True),
    }
    
def my_fit_config(trial: optuna.Trial):
    return {'dropna':False
    }
    
def my_init_config(trial: optuna.Trial):
    return {
        'lags': [24 * i for i in range(1, 3)],  # this won't be tuned
        'lag_transforms' : {
    1: [ExponentiallyWeightedMean(alpha=0.3),
        RollingMean(window_size=7, min_samples=1),
         RollingMean(window_size=21, min_samples=1),
          RollingMean(window_size=28, min_samples=1),
           RollingMean(window_size=50, min_samples=1),
        ],
}}

my_lgb = AutoModel(
    model=lgb.LGBMRegressor(),
    config=my_lgb_config,
)

import time

# -------------------------------
# Fitting model WITHOUT reusing CV splits
# -------------------------------
t0 = time.perf_counter()

auto_mlf_no_reuse = AutoMLForecast(
    models={"my_lgb": my_lgb},
    freq="D",
    season_length=24,
    fit_config=my_fit_config,
    init_config=my_init_config,
    reuse_cv_splits=False,
).fit(
    series,
    n_windows=15,
    h=horizon,
    num_samples=10,
)

t_no_reuse = time.perf_counter() - t0
print(f"Fit time without CV split reuse: {t_no_reuse:.2f} seconds")

# -------------------------------
# Fitting model WITH reusing CV splits
# -------------------------------
t0 = time.perf_counter()

auto_mlf_reuse = AutoMLForecast(
    models={"my_lgb": my_lgb},
    freq="D",
    season_length=24,
    fit_config=my_fit_config,
    init_config=my_init_config,
    reuse_cv_splits=True,   # <-- important
).fit(
    series,
    n_windows=15,
    h=horizon,
    num_samples=10,
)

t_reuse = time.perf_counter() - t0
print(f"Fit time with CV split reuse: {t_reuse:.2f} seconds")

# -------------------------------
# Comparison
# -------------------------------
speedup = t_no_reuse / t_reuse
delta = t_no_reuse - t_reuse

print(f"Speedup: {speedup:.2f}×")
print(f"Time saved: {delta:.2f} seconds ({delta / t_no_reuse * 100:.1f}%)")

Description

Solves #538

Checklist:

• This PR has a meaningful title and a clear description.
• The tests pass.
• All linting tasks pass.
• The notebooks are clean.

[janrth]

@nasaul Here is an implementation, which re-uses the CV splits to increase tuning speed.

[codspeed-hq]

Merging this PR will not alter performance

Summary

✅ 12 untouched benchmarks

---

_{Comparing janrth:enhancement/reuse_backtest_splits (c617ffc) with main (48f7d03)}

[nasaul]

Good implementation @janrth! Just minor comments to solve before merging.

[nasaul]

LGTM