Make staged primitive rules possible

[willtebbutt]

This will put the infrastructure in place to make it possible to resolve (the known problems in) #427 . It will also make it possible to move a range of other computations from rule-runtime to rule-compile time, which will reduce our dependence on constant propagation / type inference misuse, and reduce our use of generated functions further.

Before merging, I'm going to add docs.

Related (?): #403

[github-actions]

Performance Ratio:
Ratio of time to compute gradient and time to compute function.
Warning: results are very approximate! See here for more context.

┌────────────────────────────┬──────────┬─────────┬─────────────┬─────────┐
│                      Label │ Mooncake │  Zygote │ ReverseDiff │  Enzyme │
│                     String │   String │  String │      String │  String │
├────────────────────────────┼──────────┼─────────┼─────────────┼─────────┤
│                   sum_1000 │     87.0 │     1.1 │        5.51 │    8.31 │
│                  _sum_1000 │     6.61 │  1340.0 │        33.8 │    1.09 │
│               sum_sin_1000 │     2.26 │    1.67 │        10.5 │    1.98 │
│              _sum_sin_1000 │     2.68 │   241.0 │        13.3 │    2.45 │
│                   kron_sum │     62.4 │    3.51 │       223.0 │    7.98 │
│              kron_view_sum │     21.8 │     3.2 │        76.0 │    33.7 │
│      naive_map_sin_cos_exp │     2.52 │ missing │        7.56 │    2.33 │
│            map_sin_cos_exp │      2.8 │    1.46 │        6.12 │    2.91 │
│      broadcast_sin_cos_exp │     2.55 │    2.27 │        1.48 │    2.26 │
│                 simple_mlp │     7.43 │    3.12 │        12.2 │     3.6 │
│                     gp_lml │     13.2 │    6.98 │     missing │    8.79 │
│ turing_broadcast_benchmark │     3.18 │ missing │        25.6 │ missing │
│         large_single_block │     4.53 │  4240.0 │        32.0 │    2.24 │
└────────────────────────────┴──────────┴─────────┴─────────────┴─────────┘

[codecov]

Codecov Report

All modified and coverable lines are covered by tests ✅

Files with missing lines	Coverage Δ
src/Mooncake.jl	100.00% <100.00%> (ø)
src/interpreter/s2s_reverse_mode_ad.jl	94.31% <100.00%> (+0.03%)	⬆️