Make ThinkBooster installable without patches

[smirnovlad]

Summary

Step towards releasing ThinkBooster as a PyPI package. After merging lm-polygraph#446, lm-polygraph no longer needs post-install patching.

• lm-polygraph as direct dependency — added to pyproject.toml, no more sed hacks in setup.sh
• vLLM moved to main dependencies — it's the primary inference backend, not optional
• Use bundled latex2sympy — switched imports from latex2sympy2 pip package to bundled llm_tts.evaluation.latex2sympy, avoiding antlr4==4.7.2 vs 4.9.x conflict with hydra
• setup.sh cleaned up — removed all sed patches, switched from dev to main branch

What still needs setup.sh

• llm-uncertainty-head (GitHub-only)
• vllm-speculators (GitHub-only)
• KernelAct (git clone, not a pip package)

Test plan

• Clean conda env, pip install -e . only (no setup.sh)
• Full math500 eval with Qwen2.5-Math-7B-Instruct: exact_match 0.836
• vLLM inference works
• Bundled latex2sympy produces correct results

[smirnovlad]

Ready to merge. Summary of all changes:

Dependencies

• lm-polygraph>=0.6.0 from PyPI (was git+https dependency — PyPI blocker removed)
• vLLM moved from optional to main dependencies
• Removed unnecessary thinc>=8.3.0 override
• Loosened numpy to >=1.23.5

Imports

• Switched latex2sympy2 → bundled llm_tts.evaluation.latex2sympy (avoids antlr4/hydra conflict)
• Fixed isort: first-party imports in separate group

setup.sh

• Removed all sed patches for lm-polygraph
• Switched lm-polygraph from dev to main branch
• Removed post-install numpy/thinc/spacy pinning block

PyPI metadata

• requires-python >= 3.10 (lm-polygraph + vLLM requirement)
• Fixed placeholder author metadata
• Removed broken run-tts-eval console script (scripts/ is not a package)
• Removed non-existent config package-data paths
• Updated black/mypy target versions

Verified

• Clean pip install -e . in fresh conda env — no setup.sh needed
• All key imports pass (llm_tts, grader, parser, vllm, lm_polygraph)
• python -m build + twine check both pass