Refactor gemm

[willtebbutt]

This addresses #138 , and is going to provide a basis from which to explain #142 .

• Reimplement BLAS at a marginally higher level, improve performance, and check type stability
• Extend reimplementation to handle views into arrays

edit: I'm hoping that this will also improve our performance on some of the inter-AD benchmarks that we run, where I think the current implementation of gemm was a problem.

[github-actions]

Performance Ratio:

┌────────────────────────────┬────────┬─────────┬─────────────┬─────────┐
│                      Label │  Tapir │  Zygote │ ReverseDiff │  Enzyme │
│                     String │ String │  String │      String │  String │
├────────────────────────────┼────────┼─────────┼─────────────┼─────────┤
│                        sum │   40.8 │   0.433 │        3.02 │   0.659 │
│                       _sum │   7.66 │   480.0 │        27.8 │   0.134 │
│                   kron_sum │   82.4 │    3.68 │       212.0 │    24.4 │
│              kron_view_sum │   94.7 │    11.3 │       243.0 │    8.47 │
│      naive_map_sin_cos_exp │   4.32 │ missing │        8.81 │    2.77 │
│            map_sin_cos_exp │   4.62 │    1.78 │        7.72 │    3.43 │
│      broadcast_sin_cos_exp │   4.46 │    2.66 │        1.67 │    2.85 │
│                 simple_mlp │   5.36 │    2.47 │        5.19 │    2.37 │
│                     gp_lml │   6.26 │    2.34 │     missing │ missing │
│ turing_broadcast_benchmark │   6.57 │ missing │        34.2 │ missing │
└────────────────────────────┴────────┴─────────┴─────────────┴─────────┘