test(mlx-fused): refresh stale fused-loop test fakes and expectations

[FluffyAIcode]

Summary

Fixes 4 pre-existing failing tests in tests/backends/mlx/test_fused_specdecode.py that had gone stale against the current fused_specdecode_generate loop. Test-only change — no production code touched. (Surfaced as the "⚠️ pre-existing failures" noted in PR #146.)

Why they were failing (verified red on origin/main)

The fused loop evolved but its unit-test fakes/expectations did not:

1. Missing fake method. The loop now calls adapter.last_aux_torch_slice(...) (aux hidden states are captured lazily in MX and bridged to torch on demand), but _FakeAdapter lacked it → AttributeError in all 3 test_fused_loop_* tests.
2. Full-acceptance no longer appends a correction. On full acceptance the loop reuses block_logits[-1] as the next distribution instead of forwarding a correction token, so committed sequences shifted:
   • test_fused_loop_full_acceptance: appends == [104] (was [103, 105]).
   • test_fused_loop_stops_on_eos: the EOS token now arrives as block 2's bonus → blocks == 2 (was 1); tokens unchanged ([100,101,102,103]).
3. Lazy aux bridging. forward_block now leaves _last_aux = None and exposes the bridged aux via last_aux_torch_slice(); test_adapter_prefill_forward_commit now asserts that contract.

Changes

• Add last_aux_torch_slice to _FakeAdapter (mirrors MLXRestoredIncrementalVerifier).
• Re-derive expected token sequences / appends / blocks for the 3 loop tests against current behavior (full traces in the test comments).
• Update test_adapter_prefill_forward_commit to the lazy-aux contract (_last_aux is None; bridged via last_aux_torch_slice()).
• Add test_fused_loop_greedy_fallback_on_low_acceptance covering the low-acceptance → greedy-fallback branch (previously untested).

Testing

• ✅ pytest tests/backends/mlx/test_fused_specdecode.py → 14 passed (was 9 passed / 4 failed on main)
• Note: full-module 100% coverage for inference_engine.backends.mlx.* is enforced on the Mac integration run (the MLX-real paths can't import on the Linux CI host, per .coveragerc); this PR raises the Linux-exercisable loop coverage and fixes the red tests.