fix: address lit broadcasting and output name of right arithmetic ops

[AlessandroMiola]

What type of PR is this? (check all applicable)

• 💾 Refactor
• ✨ Feature
• 🐛 Bug Fix
• 🔧 Optimization
• 📝 Documentation
• ✅ Test
• 🐳 Other

Related issues

• Resolves [Bug]: lit does not "broadcast" as expected #853, resolves [Bug]: right-hand-arithmetic output name differs #509

Checklist

• Code follows style guide (ruff)
• Tests added
• Documented the changes

If you have comments or can explain your changes, please do so below

Opening as draft to get comments/guidance on the approach. Also, (1) Dask behaviour, (2) column name resulting from right arithmetics on Series in Polars, (3) leftover binary dunder methods need more thourough exploration on my side.

Another point, this PR should solve both of the issues in principle. I combined the two fixes because I originally thought that the second (#509) could naturally be resolved as a consequence of the first, but at the end that's not really the case (at least via the approach I followed). Perhaps I should separate the two?

Key points:

#853:
- Tweak validate_column_comparand in _pandas_like so as to reindex other via the lhs series only when possible (which is what the error message was complaining about when outputting ValueError: Length mismatch: Expected axis has 3 elements, new values have 1 elements)
- Broadcast the lhs series via maybe_broadcast_scalar_into_series when relevant (i.e. when of different length wrt the previously validated/broadcasted rhs other)

#509:
- Add optional alias parameter to reuse_series_implementation and assign it to expr._output_names (basically, the approach Marco was showing in last week's livestream) and pass it all along to be able to rename the output of rarithmetic ops as "literal" without incurring into safety assertion errors.

Points that need - for sure - further exploration:

• Dask behaviour related to both [Bug]: lit does not "broadcast" as expected #853 and [Bug]: right-hand-arithmetic output name differs #509 (I had to xfail tests/frame/lit_test.py::test_lit_operation and tests/expr_and_series/arithmetic_test.py::test_right_arithmetic_expr); what happens here is that Dask's lit reasonably aligns the index with the native frame, but here a reduction takes place, which shrinks the index of the resulting frame causing a mismatch.
• Right arithmetic operations on Series in Polars return an unnamed column. So, while

@nw.narwhalify
def func(df):
    return df.select(1 + nw.col('a'))

outputs "literal" as col name as reported in #509, 1 + pl.Series([1, 2, 3]) returns an unnamed series. This does prevent tests/expr_and_series/arithmetic_test.py::test_right_arithmetic_series from testing for the output column name being "literal". Point is that, as we're generally reusing series implementation in expressions, this might need to be taken into consideration as on other backends we're kind of imposing output name to be "literal".

• Analyse behaviour of the binary dunder methods that were momentaneously left apart.

[MarcoGorelli]

thanks a tonne @AlessandroMiola for starting this! I just took it a bit further, hopefully we may finally be able to resolve the linked issues 🤞

this also helped find a small bug in Polars pola-rs/polars#20071

[FBruzzesi]

I believe this should be version (1, 16, 1) ? I am fixing on another PR is that's ok

[MarcoGorelli]

thanks!