Add Entroly to Inference Optimization references

[juyterman1000]

What this PR adds

Adds Entroly to the Inference Optimization references section (§6).

Entroly is a local context compression engine for AI coding agents that reduces input tokens by 70–95% using knapsack optimization, entropy scoring, and cache alignment — without accuracy loss.

Why it fits this section:

• Inference optimization is not just about compute speed — input token cost is often the dominant expense for teams using cloud LLM APIs
• Entroly complements existing references (Flash Attention for compute, speculative decoding for latency) by addressing the input cost dimension
• All benchmark claims are backed by committed JSON artifacts and reproducible with a single command: entroly verify-claims

Key stats (committed artifacts):

• 70–95% input token reduction on large repos
• 100% accuracy retention on NeedleInAHaystack, BFCL benchmarks
• Works with 38+ AI coding tools (Claude, Cursor, Codex, Aider, etc.)

Apache-2.0, local-first, no outbound analytics by default.