Block interface: extract mixer and mlp config

[jlamypoirier]

✨ Description

Replace inheritance with composition for MLP and mixer configs

Also:

• Extract MoE config into a separate dynamic MLP type.
• Remove most ad-hoc initialization arguments
• Merge mlp_lr_scale and router_lr_scale into a single lr_scale
• Add set_defaults for defaults that depend on the hidden size

Todo (next prs):

• Add back fine-grained lr_scale
• Add back fine-grained bias enabling config (quen2 and dream disabled).
• Add back fine-grained initialization configuration.
• Rework SSM conversion (disabled).