handle finetuning errors gracefully

[arnavsinghvi11]

small change to ensure if finetuning in BootstrapFinetune fails for whatever reason, the error is raised instead of being saved and returned as the optimized program.

In current behavior, running this

optimizer = dspy.BootstrapFinetune(num_threads=16)  # if you *do* have labels, pass metric=your_metric here!
classify_ft = optimizer.compile(student_classify, teacher=teacher_classify, trainset=unlabeled_trainset)

classify_ft.get_lm().launch()

shows the error during compilation (ERROR dspy.clients.lm: Failed to ....) but doesn't actually exit and sets the student program's lm to be the error

so classify_ft.get_lm() is the error trace and launching is when the exception is raised.

Instead, the error should be raised earlier and shouldn't be set to the returned optimized program.