Fix subscript out of bounds in predict() for stepwise models with i()

[SAY-5]

Fixes #672.

predict(newdata = ...) errored with "subscript out of bounds" when the model was extracted from a stepwise estimation containing i(), e.g. reg <- feols(am ~ sw(wt + i(cyl), wt + i(cyl)), mtcars); predict(reg[[1]], newdata = mtcars).

The i() level check in predict() looped over seq_along(object$model_matrix_info) while indexing mm_info_new (the info rebuilt from newdata). For a stepwise model the stored list can hold more entries than the rebuilt one, so mm_info_new[[i]] went past the end. I bound the loop to the entries present in both lists, which keeps the existing check on new i() levels intact.

Added a regression test in the PREDICT chunk. I don't have a local toolchain to build the package's C++ here, so I couldn't run the full suite, but the change is a one-line bound on the loop and the test mirrors the existing i() prediction tests.