Inappropriate result of expr skew and kurtosis for a magic constant inputs

[JackieJin1025]

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 polars as pl
a = pl.DataFrame({"a": [1.0042855193121334] * 60})
for l in range(60):
     print(l, a.slice(0, l).select(pl.col('a').kurtosis()).item())

output

Log output

I did not obtain log here.

Issue description

expect to constantly get nan for any size, but get None when size is 0 and -2 when size >= 22

Expected behavior

import polars as pl
a = pl.DataFrame({"a": [1.0042855193121334] * 60})
for l in range(60):
      print(l, st.kurtosis(a.slice(0, l).select(pl.col('a')).to_numpy().flatten()))

get nan consistently

Installed versions

--------Version info---------
Polars:               1.5.0
Index type:           UInt32
Platform:             Linux-3.10.107-1-tlinux2_kvm_guest-0056-x86_64-with-glibc2.28
Python:               3.9.9 (main, Apr 24 2023, 09:37:21) 
[GCC 10.2.0]

----Optional dependencies----
adbc_driver_manager:  <not installed>
cloudpickle:          2.2.1
connectorx:           <not installed>
deltalake:            <not installed>
fastexcel:            <not installed>
fsspec:               2023.4.0
gevent:               22.10.2
great_tables:         <not installed>
hvplot:               <not installed>
matplotlib:           3.9.2
nest_asyncio:         1.5.6
numpy:                1.24.2
openpyxl:             3.1.2
pandas:               1.5.3
pyarrow:              15.0.2
pydantic:             1.10.18
pyiceberg:            <not installed>
sqlalchemy:           2.0.24
torch:                1.11.0+cu102
xlsx2csv:             <not installed>
xlsxwriter:           <not installed>

related issue:
#15067

[deanm0000]

hmm I guess the difference between rust's f64::EPSILON=2.220446049250313e-16 and numpy's np.finfo(np.float64).resolution=1e-15 is relevant here. I'm not sure if the right fix would be to just hard code 1e-15 instead of using f64::EPSILON, scale f64::EPSILON by the sample size (somehow), check if all inputs are mostly equal, or something else.