Fix mismatched types in Series.pow

[billylanchantin]

Fixes:

• Mismatch dtypes of Explorer.Series.pow/2 result in data frame and series when mutate with lazy series #374

New dtype handling as of Rust Polars 0.36.2:

pow(uint,  uint)  = uint
pow(uint,  sint)  = float
pow(sint,  uint)  = sint
pow(sint,  sint)  = float
pow(float, *int)  = float
pow(*int,  float) = float
pow(float, float) = float

[billylanchantin]

Hey sorry for the churn! I was playing with a version that kept everything in Elixir but I didn't mean to push it yet.

[billylanchantin]

Ok I think this new Elixir-only direction is better.

1. We're more in control of the types. @josevalim, it should now be easy to follow Nx if we want (we may need a post-cast in some cases though).
2. It helped catch some inconsistencies in other tests.
3. Less Rust code for me to mess up. 😛

[josevalim]

Smart use of DFs!

[philss]

Looks great! Nice usage of the DF, and nice way to test it! :D

But I want to document it before we merge. Otherwise anyone coming in with Polars expectations will be thrown off.

The idea here is to document for the end users, right? I'm 👍 for this.

[billylanchantin]

Thanks, y'all!

The idea here is to document for the end users, right? I'm 👍 for this.

@philss Yeah exactly. Raising on pow(unsigned, signed) is a totally sensible choice, but it differs from what Polars is doing (and Pandas I think) so I want to call it out in the docs.

[billylanchantin]

Note: I misunderstood what @josevalim was getting at with his suggestion above. He was only consolidating the cases, not making pow work like it does in Nx.

We can still do that if we want, but perhaps that should be a separate PR. If so, I think this one is good to go.

Give this comment a 👍 if we're good with that and I'll merge it. (Or someone can just merge it themselves 😄)