TimeSeriesSplit Needs a version that allows users to specify initial time

Describe the workflow you want to enable

I would like to add a new TimeSeriesSplit link called TimeSeriesRollingOriginSplit. In the current TimeSeriesSplit it is not possible for users to get an initial length.

I am happy to open a PR for this myself if it is approved.

Describe your proposed solution

This is the proposed new TimeSeriesRollingOriginSplit
class sklearn.model_selection.TimeSeriesRollingOriginSplit(initial = 5, test_size = 1, gap = 0)

initial: the initial number of samples to use. it is often common to start with a larger training period.

test_size: Used to set the test_size

gap: Number of samples to exclude from the end of each train set before the test set.

Describe alternatives you've considered, if relevant

I currently use TimeSeriesSplit in a loop, but exclude the initial splits until the train set is the size I want.

Additional context

Again. I am happy to contribute this feature.

Start with sklearn.model_selection.TimeSeriesSplit and its documented behavior, then compare it with the proposed initial, test_size, and gap parameters. Define the rolling-origin semantics and verify that the new splitter produces the requested initial training length, with corresponding documentation and coverage for its train/test indices.