[KV Cache] Support for CodeGen

[dbogunowicz]

The implementation of the KV Cache injection for the CodeGen model.
The model that was used for the injection is exported using transformers==4.30.2 (anticipating the incoming internal transformers upgrade)

Feature Preview

see: neuralmagic/deepsparse#1078

Notes:

• Decided to add a KVCacheConfig class for a clearer interface for setting up KV Cache injection parameter. I like how it makes things more structured, but I see potential disadvantages of its presence. Something I'd definitely like to talk about.
• Made CacheKeysAndValues transformation more robust and extendable.
• With this PR in place, the integration of the BLOOM model should not take more than a few hours, all the building blocks are there.
• The transformations for injecting positions input are lightweight, they share components such as add_position_input methods or a helper function for deleting orphaned nodes.

Graph Changes Post Transformers Upgrade

Codegen:

Positions

Range gets replaced with the positions input

Without KV Cache Injection

With KV Cache Injection

Attention Block

Without KV Cache Injection

With KV Cache Injection

OPT:

Attention Block

With KV Cache Injection

Without KV Cache Injection

Positions

With KV Cache Injection

("Add" node is required, see: https://github.com/huggingface/transformers/blob/main/src/transformers/models/opt/modeling_opt.py#LL105C32-L105C32)

Without KV Cache Injection

Testing Plan

• Test that the generated model gives the correct output in the pipeline (CodeGen)
• Test that the generated model gives correct output in the pipeline (OPT)
• Test the pathway, where the user provides kwargs instead of model_path to the KVCacheInjector transformation.

[KSGulin]

Looks great. Left a couple questions, but they're non-blocking

[bfineran]

looks great pending comments