feat: Support optimised init from non-dict Mapping objects in from_records and frame/series constructors

[alexander-beedie]

Added a path for from_records (and DataFrame/Series construction) that can handle conversion fromPyMapping in the Rust/pyO3 layer (currently we only handle PyDict).

This opens up efficient init from other useful Python record types that look like dicts to the caller (and support the Mapping protocol), but that we couldn't load as such (SQLAlchemy RowMapping, for example).

Also:
Fixes two long-standing errors with from_records, adding test coverage for both -

1. Incorrectly loading as Struct if the first record was a None value.
2. Failing to load at all if a None value was present after the first record:
   TypeError: 'NoneType' object cannot be converted to 'PyDict'

And:
Optimises PyDict value lookup (by precomputing the keys as PyString so we don't create them in the row-building loop).

Example

from collections.abc import Mapping
from typing import Any, Iterator

class MappingObject(Mapping):
    def __init__(self, **values: Any) -> None:
        self._data = {**values}

    def __getitem__(self, key: str) -> Any:
        return self._data[key]

    def __iter__(self) -> Iterator[str]:
        yield from self._data

    def __len__(self) -> int:
        return len(self._data)

mapping_data = [
    MappingObject(Name="Alice", Age=38, DOB=date(1987,3,5)),
    MappingObject(Name="Bob", Age=20, DOB=date(2005,5,2)),
    MappingObject(Name="Charles", Age=32, DOB=date(1993,1,18)),
] * 100_000

dict_data = [dict(d) for d in mapping_data]

Timings 🕐

(Tested with local make build-dist-release binary).

Having native conversion allows us to ingest Mapping data 2x faster than having the user convert it to dict themselves. We retain the separate dict fast-path, which remains optimal.

%timeit df = pl.from_records([dict(d) for d in mapping_data])
# 241 ms ± 2.08 ms per loop (mean ± std. dev. of 7 runs, 1 loop each)

%timeit df = pl.from_records(mapping_data)
# 124 ms ± 1.98 ms per loop (mean ± std. dev. of 7 runs, 10 loops each)

%timeit df = pl.from_records(dict_data)
# 79.7 ms ± 1.74 ms per loop (mean ± std. dev. of 7 runs, 10 loops each)

[alexander-beedie]

Looks like it needs a small update to handle None entries; will take care of it.

Update: done...

[codecov]

Codecov Report

Attention: Patch coverage is 96.90722% with 3 lines in your changes missing coverage. Please review.

Project coverage is 80.99%. Comparing base (e71569a) to head (2e01701).
Report is 23 commits behind head on main.

Files with missing lines	Patch %	Lines
crates/polars-python/src/dataframe/construction.rs	96.51%	3 Missing ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##             main   #22638      +/-   ##
==========================================
+ Coverage   80.98%   80.99%   +0.01%     
==========================================
  Files        1661     1661              
  Lines      234869   234937      +68     
  Branches     2773     2774       +1     
==========================================
+ Hits       190198   190291      +93     
+ Misses      44004    43978      -26     
- Partials      667      668       +1     

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.