Slow compile times for large kernels

Compiling a simple gemm with the following constraints takes 10 minutes.

```python
BLOCK_M: 512,
BLOCK_N: 512,
BLOCK_K: 512,

constraints: list[tkw.Constraint] = [tkw.WorkgroupConstraint(M, BLOCK_M, 0)]
constraints += [tkw.WorkgroupConstraint(N, BLOCK_N, 1)]
constraints += [tkw.TilingConstraint(K, BLOCK_K)]
constraints += [tkw.WaveConstraint(M, BLOCK_M / 2)]
constraints += [tkw.WaveConstraint(N, BLOCK_N / 2)]
constraints += [
tkw.HardwareConstraint(
threads_per_wave=64, mma_type=mfma_variant
)
]
```

I suspect there is something that does not scale well in the background which requires some investigation.

This further becomes important for our auto-tuning story, where compile times of 10s can already make the search prohibitively expensive.