Validate capacity planner accuracy and compile extensive report with next steps

[jgchn]

Description

Addresses #194

• Blog article draft: https://github.com/jgchn/llm-d-planner/blob/accuracy/accuracy/blog-gpu-capacity.md
• Included reproducibility guide for validating server deployments and analysis
• Included actual vLLM startup logs

How Has This Been Tested?

Does not affect current UI or API workflow. No effect on llm-d-planner functionality

Merge criteria:

• The commits are squashed in a cohesive manner and have meaningful messages.
• Testing instructions have been added in the PR body (for PRs involving changes that are not immediately obvious).
• The developer has manually tested the changes and verified that the changes work

Next steps

• Recalibrate activation constants in capacity planner
• Support --dtype, --kv-cache-dtype, and --quantization override in UI and API
• Fix find_possible_tps(): TP must also be divisible by vocab size. vocab_size is usually powers of 2 times a multiplier

[namasl]

When using planner, will I know if I enter some setup or model that planner doesn't fully support (will it give me some sort of error message, or proceed to give me results with large errors without warning)?