fix: scoring/output follow-ups + stop ligand recentering

[SergeiNikolenko]

Supersedes #21.

What

• finalize scoring/output correctness follow-ups on top of PR5
• align minimized scoring flow with the already accepted CLI logic from earlier PRs
• stop ligand recentering during dataset preparation

Main changes

• matcha/cli.py
  • use minimized scoring output directories consistently
  • recompute fast PoseBusters filters from minimized SDFs before top-pose selection
  • keep scoring/output selection logic aligned with the accepted CLI behavior
• matcha/dataset/pdbbind.py
  • remove ligand recentering during dataset preparation
  • preserve Dasha's small CLI/data fix already included in this branch

Why

This PR is intentionally small and functional: it makes scoring outputs consistent with the minimized flow and removes ligand recentering, which is no longer needed and should not affect current logic.

Validation

uv run pytest -n0 tests

Manual run shape to verify minimized scoring flow:

CUDA_VISIBLE_DEVICES=7 uv run matcha \
  -r /mnt/ligandpro/db/datasets/posebusters/astex_diverse_set/1N1M_A3M/1N1M_A3M_protein.pdb \
  -l /mnt/ligandpro/db/datasets/posebusters/astex_diverse_set/1N1M_A3M/1N1M_A3M_ligand_start_conf.sdf \
  --run-name pr6_scoring_validation \
  --overwrite \
  -o output \
  --scorer-path /mnt/ligandpro/soft/docking/gnina/run_gnina.sh

[SergeiNikolenko]

Moved to fork and merged into SergeiNikolenko/matcha main: https://github.com/SergeiNikolenko/matcha/tree/main