Preserve formulas in follow-up hazard-ratio refits

[dnncha]

ELI5: follow-up hazard ratios could come from a different Cox model than the one the user fitted. Formula-based fits silently expanded to every eligible dataframe column.

This forwards the saved formula when compute_followup_hazard_ratios() refits at each follow-up time.

Example

On the bundled Rossi data, a model fitted with formula="fin + age + race" returned seven terms from this method. At week 52, the reported race hazard ratio was 1.368753 instead of the fitted model's 1.300095.

Test

The regression test checks two contracts at maximum follow-up:

• returned terms equal the original formula's terms;
• hazard ratios equal the original fitted model.

python -m pytest -q lifelines/tests/test_estimation.py -k compute_followup_hazard_ratios
# 2 passed

The existing no-formula path remains covered and passes.

Fixes #1702

Limits: this fixes model specification and numerical output. It does not identify an affected publication or establish a changed scientific conclusion.