Enable ffn blocking for dense models with automatic blocking configurator

[kdulla]

Summary

• This PR introduces QEfficient ffn blocking for dense causal language models, adding token and weight blocking strategies for both MLPs with and without up projections.

Key Features

• Token Blocking: Blocked compute over hidden_states sequence length
• Weight Blocking: Blocked compute over MLP weights
• Currently, Configurable via qaic_config (explicit or auto), adjust relevant compiler flags based on ffn blocking strategy as well.
• Updated performance teams ffn blocking configurator to be calculation based instead of requiring the creation of dummy MLPs.