join_asof with strategy="nearest" joins large negative values incorrectly when -inf is present.

[Unigurd]

Checks

• I have checked that this issue has not already been reported.
• I have confirmed this bug exists on the latest version of Polars.

Reproducible example

import numpy as np
import polars as pl

pl.DataFrame({"a": [-np.inf, -(2.0**61)]}).join_asof(
    pl.DataFrame({"a": [0.0, 1.0], "b": ["lower", "higher"]}), on="a", strategy="nearest"
)


# ┌────────────┬────────┐
# │ a          ┆ b      │
# │ ---        ┆ ---    │
# │ f64        ┆ str    │
# ╞════════════╪════════╡
# │ -inf       ┆ lower  │
# │ -2.3058e18 ┆ higher │
# └────────────┴────────┘

Log output

join parallel: true
ASOF join dataframes finished

Issue description

When performing join_asof with the nearest strategy and when -inf is in the joined column in the left dataframe, values strictly smaller than -(2.0**60) don't match the nearest value in the joined column of the right dataframe.

Expected behavior

I expect -(2.0**61) to be considered nearer to 0.0 than to 1.0, and thus for the resulting dataframe to be equal to pl.DataFrame({"a": [-np.inf, -(2.0**61)], "b": "lower"})

Installed versions

--------Version info---------
Polars:              1.17.1
Index type:          UInt32
Platform:            Linux-5.15.167.4-microsoft-standard-WSL2-x86_64-with-glibc2.39
Python:              3.11.9 (main, Apr 27 2024, 21:16:11) [GCC 13.2.0]
LTS CPU:             False

----Optional dependencies----
adbc_driver_manager  <not installed>
altair               <not installed>
boto3                <not installed>
cloudpickle          3.1.0
connectorx           <not installed>
deltalake            <not installed>
fastexcel            <not installed>
fsspec               2024.10.0
gevent               <not installed>
google.auth          <not installed>
great_tables         <not installed>
matplotlib           <not installed>
nest_asyncio         1.6.0
numpy                2.0.1
openpyxl             3.1.5
pandas               2.2.3
pyarrow              18.1.0
pydantic             2.10.3
pyiceberg            <not installed>
sqlalchemy           2.0.36
torch                <not installed>
xlsx2csv             0.8.4
xlsxwriter           3.2.0

[Unigurd]

Additionally, I just noticed that

pl.DataFrame({"a": [1.0, np.nan]}).join_asof(
    pl.DataFrame({"a": [0.0, 1.0], "b": ["lower", "higher"]}), on="a", strategy="nearest"
)

yields the dataframe

┌─────┬────────┐
│ a   ┆ b      │
│ --- ┆ ---    │
│ f64 ┆ str    │
╞═════╪════════╡
│ 1.0 ┆ higher │
│ NaN ┆ higher │
└─────┴────────┘

while removing 1.0 from the left dataframe makes the computation yield

┌─────┬───────┐
│ a   ┆ b     │
│ --- ┆ ---   │
│ f64 ┆ str   │
╞═════╪═══════╡
│ NaN ┆ lower │
└─────┴───────┘

That is, whether nan is nearest to 0.0 or 1.0 depends on other values in the left dataframe.