Adds support for passing initial value to transform

[tingiskhan]

Changes made

Added support for passing a parameter called initial_value to RecursiveLinearTransform to facilitate forecasting when using the transform for latent variables.

Links to related PRs

#2037

Tests

I added one test case to the "general" part, and parameterized test_batched_recursive_linear_transform for testing arbitrary initial values.

Dependencies

No new dependencies introduced.

[tillahoffmann]

Thank you for the contribution, @tingiskhan! Do you think this could also be achieved using a composition of an AffineTransform and RecursiveLinearTransform?

[tingiskhan]

Thank you for the contribution, @tingiskhan! Do you think this could also be achieved using a composition of an AffineTransform and RecursiveLinearTransform?

That’s a very good point, I’ll take a look at it next week (if you don’t beat me to it)!

[tingiskhan]

@tillahoffmann So, I've given it some thought and see three ways forward (there might be more):

1. Keep as is (i.e. pass initial value to RecursiveLinearTransform).
2. Use a composition of AffineTransform and original RecursiveLinearTransform (as you suggested). The "issue" I see with this method is that we'll need to pass the length of the "series" on __init__ in order to construct the "loc" variable of AffineTransform correctly: loc * jnp.eye((length, 1)) (where loc corresponds to initial value). This would break the design of transforms a bit I think, since the transform will require the shape of sample on init.
3. Construct a derived class from AffineTransform that only takes the initial value as input (corresponding to loc and pass 1 to scale) and then in __call__/_inverse handle the building of the offset given the sample (same as in 2.).

Of these three I think either 1. or 3. is the most appropriate. I'm not sure how re-usable the new component of 3. would be, but I'm open to either suggestions.

I'm not sure this PR is even needed to be honest, but the current way of passing initial values other than 0 is not so clear (might just require an entry to the docs?).

[fehiepsi]

If initial value is 0, we have y_1=x_1, so we can just prepend the initial value to x and y?