fix(requirements): add missing runtime deps; pin antlr4 to 4.7.2

[ffrappo]

Problem

Out of the box, bash evaluations/install.sh + bash evaluations/verify_benchmarks.sh fails on 11 of 19 benchmarks because several runtime dependencies are missing from evaluations/requirements-main.txt. The vendored packages (lm_eval, livecodebench, lmms_eval, visualwebbench) are installed with --no-deps, so their own extras are never pulled in.

Fix

Adds four missing deps and pins antlr4 to resolve a silent three-way conflict:

dep	why
tenacity	lm_eval/models/api_models.py imports it for API-model retry. Without it, all 8 lm_eval API-model benchmarks (aime25, gpqa, ifeval, logiqa, humaneval_chat, mbpp_chat, mmlu_cot, acp_gen) raise ModuleNotFoundError: Attempted to use an API model, but the required packages ['tenacity'] are not installed.
anthropic	livecodebench/lcb_runner/runner/claude_runner.py imports Anthropic at module load even when using the OpenAI endpoint. Without it, livecodebench fails on import.
loguru	lmms_eval/tasks/__init__.py imports loguru unconditionally as eval_logger. Without it, mmmu and any lmms_eval task fail on import.
google-generativeai	visualwebbench/run.py + gemini_adapter.py import it at module load. Without it, visualwebbench fails on import.
antlr4-python3-runtime==4.7.2	tarski==0.8.2 (planbench PDDL parsing) requires ATN v3 (antlr4 4.7.x). latex2sympy2 (an optional lmms_eval math util) pulls antlr4 4.13.x (ATN v4), which silently breaks planbench plan tasks with: Could not deserialize ATN with version 3 (expected 4). Pin 4.7.2 to keep planbench working.

Verification

Fresh clone + install.sh + verify_benchmarks.sh:

Before: 8 passed, 11 failed.

After: 19/19 PASS (16 text benchmarks via a text endpoint + 3 multimodal via a vision endpoint).

===================== SUMMARY =====================
  PASS  infobench  PASS  aime25  PASS  gpqa  PASS  ifeval  PASS  logiqa
  PASS  humaneval_chat  PASS  mbpp_chat  PASS  mmlu_cot  PASS  acp_gen/acp_app_gen
  PASS  livecodebench/codegeneration  PASS  repobench/repobench_xff_python
  PASS  debugbench  PASS  lifbench/onedoc-qa  PASS  planbench/task_3_plan_verification
  PASS  mmmu  PASS  visualpuzzles/cot  PASS  visualwebbench/web_caption
  PASS  beir_nfcorpus  PASS  bfcl/non_live_parallel_multiple
  19 passed, 0 failed

Tested on Ubuntu 24.04, Python 3.11, with an OpenAI-compatible endpoint.