SpD, multiprojection heads

[eplatero97]

Objective

This PR seeks to implement post-attention hidden size projections used to speculate tokens ahead of the base model. This PR contains three primary components:

1. extending base model with multi-projection in modeling_auto.py
2. implementing multi-projections forward pass.
3. app demo of multi-projection model.

Initial Implementation

Initial implementation gives the user the flexibility to define its own projection architecture and pass it to QEffAutoModelForCausalLM.

Then, QEfficient simply attaches these projections to the model to be used during the forward pass. The "attaching" of these projections is done by using the accelerate library. I used this library because it has a robust implementation to attach weights to an already existing model. We can implement our own abstraction if needed, but first we must agree on what the external API will be to the user.

NOTE: Please keep in mind that to integrate medusa, similar changes will be needed (instead of doing multiple hidden size projections, medusa uses multiple lm_heads to speculate ahead of the base model).

[eplatero97]

using sdk 1.20.0.63, I manually validated that all qwen vanilla spd tests passed as well as llama multiprojection pytorch unit test