python: prefer pyarrow when we can memory map the file

[ritchie46]

This change prefers the pyarrow reader because we can memory map the file. This will only trigger an IO read syscall when we actually read the colums, and only the parts that we read.

In lazy we can do projection/predicate pushdown on the memory mapped columns. The columns that will be filtered out never trigger a read call.

Another advantage is that we might be able to work on larger than memory data as the OS will purge pages back that are not used when it needs more RAM.

If the files do easily fit in RAM, we can query from the file system. Subsequent queries on a file may be much faster as the file/or large parts of it can still be in RAM.

I have some ideas for being able to do this natively as well. Might follow up on that.

[codecov-commenter]

Codecov Report

Merging #4182 (6b0efa7) into master (08f6f73) will increase coverage by 0.01%.
The diff coverage is 50.00%.

@@            Coverage Diff             @@
##           master    #4182      +/-   ##
==========================================
+ Coverage   78.75%   78.77%   +0.01%     
==========================================
  Files         458      458              
  Lines       75783    75785       +2     
==========================================
+ Hits        59685    59701      +16     
+ Misses      16098    16084      -14     

Impacted Files	Coverage Δ
py-polars/polars/io.py	72.82% <50.00%> (+2.49%)	⬆️
...ars/polars-core/src/chunked_array/ops/any_value.rs	77.39% <0.00%> (-0.87%)	⬇️
...polars-time/src/chunkedarray/rolling_window/mod.rs	71.53% <0.00%> (-0.77%)	⬇️
py-polars/polars/testing.py	94.08% <0.00%> (-0.54%)	⬇️
...lars/polars-core/src/chunked_array/ops/take/mod.rs	63.12% <0.00%> (-0.34%)	⬇️
polars/polars-io/src/csv/buffer.rs	79.39% <0.00%> (-0.26%)	⬇️
polars/polars-arrow/src/kernels/take.rs	87.40% <0.00%> (+0.24%)	⬆️
polars/polars-core/src/vector_hasher.rs	77.15% <0.00%> (+0.26%)	⬆️
...olars/polars-core/src/frame/groupby/into_groups.rs	60.47% <0.00%> (+0.29%)	⬆️
...s/polars-core/src/chunked_array/ops/unique/rank.rs	96.19% <0.00%> (+0.34%)	⬆️
... and 6 more

---

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update 08f6f73...6b0efa7. Read the comment docs.