Lazy filter/validate plans contain absolute install paths causing serialized plans to differ between environments #413

My pipelines use LazyFrame.serialize() bytes to determine whether the execution plan has changed.

When using Schema.filter(..., eager=False) (also validate and cast) inside these pipelines, the lazy result contains two things that depend on where dataframely is installed. This causes the same code on the same data to produce different LazyFrame.serialize() bytes when run from a different machine/user/checked out directory, etc. Is there any way to prevent this / fix this in dataframely?

In the plan I can see:

1. An absolute path to dataframely installation: all_rules_horizontal, all_rules, all_rules_required are registered with use_abs_path=True. The plan stores e.g. /home/david/project/.pixi/envs/default/lib/python3.13/site-packages/dataframely/_native.abi3.so.

2. A Python function stored in the plan: match_to_schema passes cast_lenient/cast_none/cast_strict to lf.pipe_with_schema(...), so the function is stored in the plan. These are nested functions within match_to_schema, so serializing the plan pickles their full code, including the absolute path of dataframely/_match_to_schema.py.

Example:

import sys

import dataframely as dy
import polars as pl


class MySchema(dy.Schema):
    a = dy.Int64(nullable=False)


lf = pl.LazyFrame({"a": [1, None]})
valid, failure = MySchema.filter(lf, cast=True, eager=False)
plan = valid.serialize()

print(sys.prefix.encode() in plan)  # True: absolute env path is part of the plan
print(b"cloudpickle" in plan)  # True: a pickled Python function is part of the plan

(dataframely 3.1.2, polars 2.0.0, Python 3.13)