Release GasBench v0.11.4 with corrected audio provenance labels

[dylanuys]

Human Thorsten recordings and ASVspoof 2021 physical replay speech were labeled synthetic, while mixed ASVspoof and FakeSound2 sources were assigned a single class. This release corrects those labels and selects separate real and synthetic subsets using the existing registry filters.

• Move Thorsten-Voice and ASVspoof 2021 PA to real audio.
• Use pinned provenance-separated mirrors for ASVspoof 2021 LA/DF and ASVspoof 5; add real LA/DF entries and a synthetic neural-codec reconstruction entry for C04/C07.
• Switch FakeSound2 to its original source and separate genuine recordings, recorded edits and LASSNet masking, and generatively produced audio with archive-member filters.
• Bump the package version from 0.11.3 to 0.11.4.

Includes #186, merged into dev. The release diff changes the two audio registry files and pyproject.toml.

Validation: all four CI checks passed on #186; nine existing config-loading/archive-filter tests were rerun successfully; ruff check src tests and git diff --check passed; pyproject.toml parses successfully with version 0.11.4. Prior source validation included fresh downloads from all six corrected ASVspoof subsets matched to official protocol labels, a complete FakeSound2 class partition, archive CRC checks, and SHA-256 verification of duplicate audio files.

Sources: ASVspoof LA, ASVspoof DF, ASVspoof 5, FakeSound2.