Improve Small-Union Performance

[willtebbutt]

Improves the way that unions are handled to mean that the performance of small unions is substantially improved.

There's still an allocation associated to the forwards-pass of pi nodes involving small unions that I've not yet managed to track down, but the performance is otherwise quite good now.

[github-actions]

Performance Ratio:

┌────────────┬────────┬─────────┬──────────────┬─────────┐
│      label │  taped │  zygote │ reverse_diff │  enzyme │
│     String │ String │  String │       String │  String │
├────────────┼────────┼─────────┼──────────────┼─────────┤
│        sum │ 2850.0 │    1.14 │         7.92 │    1.84 │
│       _sum │  426.0 │  1220.0 │         71.3 │   0.307 │
│       kron │ 1490.0 │    3.51 │        209.0 │    27.3 │
│ kron types │ 1600.0 │    11.1 │        326.0 │    8.62 │
│  naive map │   36.9 │ missing │         9.16 │    2.81 │
│        map │   55.8 │    1.76 │         7.75 │    3.48 │
│  broadcast │   56.0 │    2.69 │          1.7 │    2.87 │
│        mlp │  118.0 │    3.11 │         11.2 │    3.64 │
│         gp │  335.0 │    4.09 │      missing │ missing │
└────────────┴────────┴─────────┴──────────────┴─────────┘