Abstract Fields in Arguments

[willtebbutt]

Addresses #184

[github-actions]

Performance Ratio:
Warning: results are very approximate!

┌────────────────────────────┬────────┬─────────┬─────────────┬─────────┐
│                      Label │  Tapir │  Zygote │ ReverseDiff │  Enzyme │
│                     String │ String │  String │      String │  String │
├────────────────────────────┼────────┼─────────┼─────────────┼─────────┤
│                        sum │   35.0 │   0.324 │        2.35 │   0.486 │
│                       _sum │   6.76 │   505.0 │        26.6 │   0.121 │
│                   kron_sum │   85.0 │    3.51 │       219.0 │    24.1 │
│              kron_view_sum │   97.8 │    11.4 │       255.0 │    8.66 │
│      naive_map_sin_cos_exp │   4.12 │ missing │        8.98 │     2.8 │
│            map_sin_cos_exp │   4.76 │    1.79 │        7.56 │     3.4 │
│      broadcast_sin_cos_exp │   4.49 │    2.66 │        1.69 │    2.86 │
│                 simple_mlp │   8.79 │    3.09 │        11.3 │    3.15 │
│                     gp_lml │   15.2 │    4.31 │     missing │ missing │
│ turing_broadcast_benchmark │   8.91 │ missing │        28.7 │ missing │
└────────────────────────────┴────────┴─────────┴─────────────┴─────────┘

[willtebbutt]

Hope fully the latest commits fix the issue with the Turing.jl integration testing. This should not be merged yet though, because I need to add a unit test which captures the behaviour that was causing the Turing.jl integration test to fail.

It's mildly awkward to add a unit test for this, and this code is unlikely to regress. If we find the problem that I found half way through this PR in a future PR, we'll add a unit test there. I'll merge when CI passes.