feat: compact TOON encoder, response caching, optional Semgrep pre-pass

[tusharshah21]

• Rewrote TOON as a tabular format (F:/C[N]{op,ln,code}: schema + CSV rows). ~45% smaller than the old per-chunk JSON, ~20% smaller than raw unified diff on a representative two-file diff. Context lines around changes are trimmed (configurable via CONTEXT_LINES, default 2).
• Encoder now operates at the file level so F: header is emitted once per file.
• Added LLM response cache keyed by SHA256(model, messages). Skips re-billing duplicate chunks within a job; toggle via ENABLE_CACHE (default on).
• Optional Semgrep pre-pass: when SEMGREP_RULES is set and semgrep is on PATH, findings are passed to Agent 1 as priors. Agent 1 still verifies before flagging to filter false positives.
• Fixer (Agent 2) now reads +/-15 lines of surrounding file context from the checked-out workspace when available, giving it imports and enclosing signatures for better fix accuracy.
• README: documented Ollama and GitHub Models as new free-provider options, added Semgrep section, replaced aspirational "50-70%" line with benchmarked numbers, documented new inputs.
• Bumped to 1.1.0 with CHANGELOG.