[Tracing] Trace with eager attention

[kylesayrs]

Purpose

• Enable higher-granularity tracing for llama-style models
• Tracing at the linear level requires tracing into the attention mechanism, which can be complex for non-eager implementations

Changes

• Use eager attention implementation when tracing

Testing

• After [Tracing] Allow torch.Sizes to be iterated #1308 lands, you can now perform sequential calibration of llama3 models with Linear granularity

[github-actions]

👋 Hi! Thank you for contributing to llm-compressor. Please add the ready label when the PR is ready for review.

Note: This is required to complete the testing suite, please only add the label once the PR is code complete and local testing has been performed.