Add cacheflow prompt-cache simulator

[ron2111]

Summary

Adds a new dependency-free cacheflow package for token-aware prompt-cache planning and trace simulation.

This intentionally moves cache APIs out of the root omnitoken package because cache simulation can grow independently from the tokenizer core.

What changed

• Added cacheflow/ package.
• Moved single-prompt boundary planning into cacheflow.NewAligner / cacheflow.AlignTokenCount.
• Added JSONL prompt trace support with cacheflow.ReadJSONL.
• Added stable token-prefix simulation with cacheflow.Simulate.
• Added best-effort cache-breaker hints for timestamps, UUID/request IDs, and dynamic JSON metadata.
• Added omni cache-sim CLI.
• Updated omni cache to use cacheflow.
• Added cacheflow/README.md and linked it from README/CLI docs.
• Removed old root cache API and docs/cache.md per the requested no-backward-compatibility cleanup.

Issues

Addresses the design and MVP direction for:

• Design: Prompt Cache Simulator for LLM Workloads #1
• Design: Define cache-sim input format #5
• Feature: Detect Likely Cache Breakers #6
• Docs: Prompt Caching Guide #4
• partially Benchmark: Add Long Prompt and Cache Simulation Fixtures #8

Verification

• go test ./...
• go vet ./...
• adapter module tests for Gemini, Hugging Face, Llama3, Mistral, OSS, and Anthropic
• CLI smoke tests:
  • omni cache
  • omni cache-sim -input prompts.jsonl -breakers -json

Notes

No new dependencies were added. Cacheflow remains a local planning/simulation tool and does not claim provider billing parity.