[WIP] TensorResult abstraction

[ivarflakstad]

Currently when we chain math with Result<Tensor> we have to write code like

(&(&(&x.powf(2.0)? * 3.0)? - &x * 4.0)? + 5.0)?

Given that we change all our methods to return the newtype TensorResult instead of Result<Tensor>, we could simply write

(x.powf(2.0) * 3.0 - x * 4.0 + 5.0).ok()?;

Where .ok() does TensorResult -> Result<Tensor>

The only unanswered question is what to do when we have two Err results.
My proposed solution is to simply aggregate them. Either via a new candle_core::Error enum type that wraps two, or by simply using Error::context (e1.context(e2)).

Ref #691

[EricLBuehler]

Looks awesome.

The only unanswered question is what to do when we have two Err results.
My proposed solution is to simply aggregate them. Either via a new candle_core::Error enum type that wraps two, or by simply using Error::context (e1.context(e2)).

I think a new Error variant would be better, since the .context might not exactly match expectations.

Maybe this can tie in with some of our specialized backend work?

[sempervictus]

Big fan of explicit error-typing for maintainability and security reasons. Convenience tucks of types into others can later be abused for type-confusion logic by red team even though Rust itself is thought to be safe from such things.