[Observability] Provide analysis scripts to extract and plot relevant per-component metrics

[raduioanstoica]

Problem: There's no built-in way to visualize many low-level metrics (KV-cache hit rate per tier, I/O time, or GPU memory occupancy) from a completed run.

Desired functionality:

• Provide post-simulation plotting/analysis tooling for all key performance metrics. Essentially, provide enough data to answer the "where does time go?" question.
• Provide a way to batch-generate all plots for a given simulation run's output directory.