Port SEQTaRget#173 fixes, plus further weighting and output fixes

[remlapmot]

Ports the subset of fixes that were needed from CausalInference/SEQTaRget#173:

• numeric fixed covariates stay continuous, weights no longer depend on the order of treatment_level,
• excused-censoring weights and the excused-column direction are corrected,
• each subject's first row is left out of the pre-expansion weight models,
• subgroup bootstrap results are matched by name,
• the end-of-follow-up eligible count is corrected, and
• undefined model terms such as np.log(x) give an informative error.

Also fixes

• the LTFU censoring weights (now P(remaining), as in SEQTaRget),
• centred splines with glum offloading
• bootstrap predictions with stateful terms
• followup_spline with km_curves
• collect() for subgroup analyses
• weight_p99 overwriting the supplied bounds.

And

• Adds Python 3.15 and tests on that.

Several of these change estimates, so have updated a few expected test values and the relevant vignette results.

[ryan-odea]

Looks good Tom, thanks! Might have to borrow your workaround for 3.15 inclusion in some other packages.