Simple speedups to decomposition

[danbraunai-goodfire]

bfloat16 everywhere is ~3x faster, but the runs seem to fail:

• run. bf16. ss_llama_simple_mlp-1L (same config as here)
• run. bf16. ss_llama_simple_mlp-1.25M (same config as here)

Mixed precision seems to be about 50% faster. Runs pending to check that it works well:

• run. 1L mixed_precision
• run. 1.25M mixed precision

We should at least use mixed precision if it works well.

We should also:

• torch.compile() the target model (compiling other things didn't seem to help/weren't possible)
• Use regular torch ops rather than einops inside the slow parts (maybe just LinearComponents.forward()).
• Turning off use_delta_components. It's 13% faster without them