FIX(py/twine): decoder cache leaks across artifacts

[hon-gyu]

In one of downstream library's tests, the twine decoding gave different results when the test is run isolated vs. in a test suite. It reproduced only with enough prior artifact decodes in the same process to populate the colliding offset.

Summary

twine.Decoder.caches is declared as a class-level mutable dict and is never
re-initialized in __init__, so every Decoder instance shares the same cache
object. Because cached values are keyed by byte-offset within a single artifact's
buffer, offsets from one artifact collide with offsets from a previously-decoded
artifact, and the decoder returns stale decoded values that belong to a different
artifact.

Reproduction

from imandrax_api.lib import twine

d1 = twine.Decoder(b"\x00")
d2 = twine.Decoder(b"\x00")
assert d1.caches is d2.caches           # BUG: same object

d1.caches.setdefault("T", {})[42] = "STALE"
print(d2.caches["T"][42])   # -> "STALE": d2 sees d1's cache

Fix

init cache per Decoder instance (per artifact)

[wintersteiger]

🤦
LGTM, thanks for the fix! Perhaps we should just remove this cache, and perhaps some others too... users will cache their artifact by themselves if they think there is a need.