Implement `compile_decoder_for_dem` for best performance from sinter

[Strilanc]

The method that decodes from disk works, but it is much worse than the one that first configures the decoder and then uses it many times so it can work in memory.

Example: https://github.com/quantumlib/Stim/blob/main/glue/sample/src/sinter/_decoding_pymatching.py#L24