Fix integerish conversions

[Ilia-Kosenkov]

Fixes #720.

• Rework scalar f64 (Rfloat) to integer type conversions (both FromRobj and TryFrom) by extracting all logic into a single FloatToInt trait.
• Properly handle all edge cases like NA, over-/under-flow, subnormality
• Disallow rounding casts that result in loss of accuracy (e.g., 4.2f64 no longer maps to 4i32)
• Allow only f64 values that are exactly representable by target integer type

Also update xtask:

• cargo extendr devtools-test now gained -f/--filter option that controls which tests to execute. Value is passed verbatim to filter parameter of devtools::test()

[CGMossa]

Lgtm.

Personally, I didn't like the way it was done before.

We should strive for a performant conversion, if you ask me, but that's another discussion altogether.
I believe if the num crate does this in a more clever way, we ought to consider it.

Bytemuck does not do conversions. I checked. It merely reinterprets bytes.

[CGMossa]

Please consider this https://doc.rust-lang.org/std/primitive.f64.html#method.to_int_unchecked
Before merging.

[Ilia-Kosenkov]

It practically does what f64 as i32 does, with perhaps tiny variations. It still will overflow/underflow to max/min value (what we observed in the original issue), it will break on various NaN, it will not propagate NA, which we definitely want. So it is practically useless to us. We are looking for a _checked version, which is not yet present.

Since I extracted the logic as much as I could, if we find a better solution to map an f64 into i32 | ConversionError, it is now easy to implement, replacing code in one place.