feat: MLX memory optimizations — bf16, fused decode, deterministic GC
Phase 1: selective bf16 mixed precision (backbone fp32, decoders bf16,
sigmoid always fp32). dtype param on GreenFormer, all outputs guaranteed
fp32.
Phase 2: FusedDecoderPair batches 3 upsample ops by concatenating
alpha+fg projections along channel axis. Bit-exact parity with unfused
path.
Phase 3: deterministic GC pipeline in tile loop (del + gc.collect +
mx.clear_cache) and engine.py intermediate tensor cleanup.
7 new tests: bf16 dtype/range/nan, fused-vs-unfused parity, fp32
default unchanged. 80 passed, 0 failures.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
b3b7df271d