Fix interface

[willtebbutt]

Fortunately, #97 was a bug in the interface, not the underlying functionality.

I forgot to the copy the output of the forwards-pass before running the reverse-pass, which is necessary in general because the reverse-pass restores the state of data to the state it was initialised to. For arrays, this is whatever it was allocated as, which is non-deterministic.

Anyway @gdalle this should fix your problem. Will merge when CI passes.

[github-actions]

Performance Ratio:

┌────────────────────────────┬────────┬─────────┬─────────────┬─────────┐
│                      Label │  Taped │  Zygote │ ReverseDiff │  Enzyme │
│                     String │ String │  String │      String │  String │
├────────────────────────────┼────────┼─────────┼─────────────┼─────────┤
│                        sum │  125.0 │   0.403 │        2.76 │   0.664 │
│                       _sum │   26.4 │   481.0 │        27.9 │   0.136 │
│                   kron_sum │  196.0 │    3.55 │       196.0 │    23.1 │
│              kron_view_sum │  238.0 │    10.7 │       226.0 │     8.0 │
│      naive_map_sin_cos_exp │   7.91 │ missing │        8.02 │    2.53 │
│            map_sin_cos_exp │   10.8 │    1.71 │        7.83 │     3.4 │
│      broadcast_sin_cos_exp │   9.96 │    2.58 │        1.65 │    2.81 │
│                 simple_mlp │   24.1 │    3.04 │        11.6 │    3.84 │
│                     gp_lml │   43.9 │    4.38 │     missing │ missing │
│ turing_broadcast_benchmark │   19.5 │ missing │        35.8 │ missing │
└────────────────────────────┴────────┴─────────┴─────────────┴─────────┘