Add python scripts for DEM manipulation: generalize_dem.py and decomposer.py

[noajshu]

Here we open source python versions of two DEM processing utilities.

• DEM Generalization: When decoding the color code experiments, we found a large improvement in logical error rate (about 0.87x reduction factor) from calibrating a DEM on the rate that a most-likely-error decoder predicts that each error mechanism across a calibration dataset taken with the same set of qubits. To aggregate stats and generalize a DEM from, say, 10 rounds to 100 rounds, we use the detector coordinates of the DEM to identify a unique spatial key.

• DEM Error Decomposition: Another DEM processing task important for matching decoders is decomposition. Stim's decomposition tactics are robust and general, at the cost of being suboptimal for the surface code.

As a next step, we could expose the functionality in Tesseract as follows:

import tesseract_decoder

# Take one or more DEMs **with detector coordinates**, aggregate the error probabilities
nice_calibrated_dem = tesseract_decoder.demutil.regeneralize_spatial_dem(
    templates: List[stim.DetectorErrorModel],
    scaffold: stim.DetectorErrorModel,
    verbose: bool=False
)

# Re-decompose the errors assuming Stim surface code coordinate convention
nice_matchable_dem = tesseract_decoder.demutil.decompose_errors(dem method='stim-surfacecode-coords')

# Re-decompose the errors assuming the last-coordinate index indicates the component:
nice_matchable_dem2 = tesseract_decoder.demutil.decompose_errors(dem, method='last-coordinate-index')

[LalehB]

LGTM!
p.s. we will move them later for demutil