enh: allow for scalars in sum_horizontal

[MarcoGorelli]

e.g.

import duckdb
import narwhals as nw
rel = duckdb.sql('select * from values (1, 4), (1, 2), (2, 3), (2, 4) df(a, b)')
df = nw.from_native(rel)
 df.select(nw.sum_horizontal(1, 'a', 'b'))

throws

InvalidIntoExprError: Expected an object which can be converted into an expression, got <class 'int'>

Hint:
- if you were trying to select a column which does not have a string
  column name, then you should explicitly use `nw.col`.
  For example, `df.select(nw.col(0))` if you have a column named `0`.
- if you were trying to create a new literal column, then you
  should explicitly use `nw.lit`.
  For example, `df.select(nw.lit(0))` if you want to create a new
  column with literal value `0`.

I think we should have a way of allowing for anything which can't be parsed as an expression to be parsed as nw.lit in these cases

[FBruzzesi]

Are we aiming for the same behavior for boolean's (e.g. in any_horizontal and all_horizontal)?

Additionally, polars allow to pass numbers to (any|all)_horizontal and booleans to (sum|mean|max|min)_horizontal. Thoughts on this?

I have a way to address it, which is failing for dask only (I believe due to some broadcasting issues for how we create the lit)

[MarcoGorelli]

thanks for looking into it!

not sure, will think about it, but this probably is quite low-prio for now, and may require a bit of a refactor if we want to do it properly (and not just special-case a few functions)

[FBruzzesi]

this probably is quite low-prio for now

I agree, using nw.lit(value) should work anyway, and that's what the error message suggest to do.

may require a bit of a refactor if we want to do it properly (and not just special-case a few functions)

I was able to tackle it by adding a few extra checks/cases in extract_compliant function, however for different methods and backend we might require casting to the correct type