audiolabs/DEISM: Performance optimization (shoebox)

Shoebox Numba image generation now has finer parallel work splitting by parity and reflection order. Shoebox v2-numba image generation supports a compact image format. The compact format stores only shoebox reflection metadata and geometry: A R_sI_r_all R_s_rI_all R_r_sI_all The compact format does not materialize atten_all, atten_all_early, or atten_all_late during image generation. Shoebox compact storage is now supported for both RIR and RTF. For shoebox v2-numba, compact storage is the default unless explicitly disabled with shoeboxCompactImages = False. merge_images() now supports both materialized and compact shoebox images. Shoebox Numba ORG, LC, and MIX now reconstruct attenuation on the fly from compact metadata. Shoebox Numba ORG, LC, and MIX all run in bounded image batches. The existing numParaImages limit is still respected. Extra temporary-memory caps are available through: numbaORGTempTargetMB numbaLCTempTargetMB The shoebox Ray backend does not support compact images and now fails with a clear message instead of silently misbehaving.