Fix zero multimodal learning rates in Megatron

[taking-lying-flat]

Summary

• honor explicit vit_lr=0.0 and aligner_lr=0.0 values in Megatron parameter groups
• assign concrete max_lr and min_lr values directly instead of representing multimodal overrides as ratios of the LLM learning rate
• preserve generic learning-rate scaling and decoupled learning-rate behavior

Root cause

The multimodal parameter-group logic checked vit_lr and aligner_lr by truthiness. Explicit zero values therefore fell through to the default LLM learning rate instead of freezing those modules. Nonzero overrides were also converted to override / lr, which could divide by zero when the base learning rate was zero.

Impact

Users can safely set either multimodal learning rate to zero, and configurations with a zero base learning rate no longer encounter division by zero while constructing optimizer groups.