projection_pushdown regression/error from >= 0.14.15 (changes types, causes SchemaMismatch Error)

[alexander-beedie]

Polars version checks

• I have checked that this issue has not already been reported.

• I have confirmed this bug exists on the latest version of Polars.

Issue description

Projection pushdown can unexpectedly change types/values.
The below example succeeds in 0.14.14, but raises an exception in all versions since.

Reproducible example

Setup:

import polars as pl
a, b, c = map( pl.col, 'abc' )

x = pl.DataFrame(
    data = [
        ('a', 1, 0),
        ('a', 0, 1),
        ('a', 1, 2),
        ('b', 0, 0),
    ],
    columns = ['a','b','c'],
).lazy()

y = x.select( [a, b.take(c).cumsum().over(a) > 0] )
z = y.select( [a, (b ^ b.shift(1).over(a)).fill_null(True).alias('keep')] )

Error: (with/without projection_pushdown)

print( z.collect(projection_pushdown=False) )
# ┌─────┬───────┐
# │ a   ┆ keep  │
# │ --- ┆ ---   │
# │ str ┆ bool  │
# ╞═════╪═══════╡
# │ a   ┆ true  │
# │ a   ┆ false │
# │ a   ┆ false │
# │ b   ┆ true  │
# └─────┴───────┘

print( z.collect(projection_pushdown=True) )
# ┌─────┬──────┐
# │ a   ┆ keep │
# │ --- ┆ ---  │
# │ str ┆ i64  │
# ╞═════╪══════╡
# │ a   ┆ 1    │
# │ a   ┆ 1    │
# │ a   ┆ 1    │
# │ b   ┆ 1    │
# └─────┴──────┘

z.filter( pl.col('keep') ).collect()
# ComputeError: Filter predicate must be of type Boolean,
#  got: SchemaMisMatch(Owned("Series of dtype: Int64 != Boolean"))

Expected behavior

Projection pushdown shouldn't result in incorrect types/values.
Expected result:

# ┌─────┬──────┐
# │ a   ┆ keep │
# │ --- ┆ ---  │
# │ str ┆ bool │
# ╞═════╪══════╡
# │ a   ┆ true │
# │ b   ┆ true │
# └─────┴──────┘

Installed versions

---Version info---
Polars: 0.14.17
Index type: UInt32
Platform: macOS-12.6-arm64-arm-64bit
Python: 3.9.7 (default, Aug 12 2022, 09:02:06) 
[Clang 13.1.6 (clang-1316.0.21.2.5)]

---Optional dependencies---
pyarrow: 9.0.0
pandas: 1.4.4
numpy: 1.23.3
fsspec: 2022.7.1
connectorx: 0.3.0
xlsx2csv: <not installed>
matplotlib: 3.5.3