Fix init_value and end_value in cosine decay

[copybara-service]

Fix init_value and end_value in cosine decay

Problem:

• the starting value was not always init_value
• the last value was not always end_value

Solution:

• changed the formulas (they are now compatible with the pytorch implementation https://pytorch.org/docs/stable/generated/torch.optim.lr_scheduler.CosineAnnealingLR.html)
• added test to check that starting and end value coincide with init_value and end_value respectively

Misc:
renamed alpha -> end_value in warmup_cosine_decay_schedule