chore: remove unused hydra-core dependency

[SergeiNikolenko]

Summary

• remove hydra-core from runtime dependencies in pyproject.toml
• keep the cleanup minimal and limited to dependencies required for Matcha inference/CLI

Why

hydra-core is still listed as a runtime dependency on main, but there are no tracked imports or references to Hydra in the current codebase. Matcha uses omegaconf directly for configuration loading/merging in the CLI and scripts.

Validation

• git grep -n 'hydra' only matched pyproject.toml
• local: uv run python -c "import matcha; print(\"import ok\")"
• local: uv run matcha --help
• local: uv run pytest tests/test_multigpu_cmd_build.py -q
• lobachevsky: fresh uv sync in a clean checkout without hydra-core
• lobachevsky: uv run matcha --help

Notes on server smoke run

I also attempted a full smoke inference run on lobachevsky using the standard 7EW9/05C Matcha smoke inputs.

Two unrelated blockers showed up:

1. Fresh uv sync currently resolves torch 2.11.0+cu130, which does not work with the server driver (found version 12080), so CUDA is unavailable in the fresh environment.
2. Forcing the run through the older working GPU environment reaches the real pipeline, but then fails with an existing unrelated runtime error:
   AttributeError: 'PDBBind' object has no attribute 'dataset_name'

Neither issue is caused by this dependency cleanup.

[SergeiNikolenko]

Superseded by #45, which now includes the same direct dependency cleanup (hydra-core) together with the CI/PyPI packaging changes.