Added wrapping_cast to cast kernels

[sundy-li]

fixes #252

[sundy-li]

It's not backwards compatible, but it follows the ways as the modern number cast did.

[codecov]

Codecov Report

Merging #254 (647fddb) into main (e92f47a) will increase coverage by 0.03%.
The diff coverage is 57.64%.

@@            Coverage Diff             @@
##             main     #254      +/-   ##
==========================================
+ Coverage   77.31%   77.35%   +0.03%     
==========================================
  Files         232      232              
  Lines       19949    20094     +145     
==========================================
+ Hits        15424    15543     +119     
- Misses       4525     4551      +26     

Impacted Files	Coverage Δ
src/compute/cast/mod.rs	43.42% <56.50%> (+8.27%)	⬆️
src/compute/cast/dictionary_to.rs	47.61% <66.66%> (ø)
src/compute/cast/primitive_to.rs	58.42% <100.00%> (+2.47%)	⬆️

---

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update e92f47a...647fddb. Read the comment docs.

[jorgecarleitao]

Makes sense. Thanks a lot. 4 observations:

1. could you point me to somewhere describing "the ways as the modern number cast did."? (I feel I should read on that :))
2. could we add a test for other cases (I think we need 4 csaes: (ignore_overflow = true and not overflowing, ignore_overflow = true and overflowing, etc.)
3. For all cases where we up-cast (e.g. i32 -> i64), shouldn't we ignore the option and always hit the fast path since it is guaranteed to not overflow?
4. I wonder if the cast option could somehow follow an equivalent notation as Rust's own notation. If I understand correctly, ignore_overflow is equivalent to using as, which is equivalent to a saturating the cast

[sundy-li]

could you point me to somewhere describing "the ways as the modern number cast did."

Emm, this is just c-style cast or default cast behavior in MySQL/ClickHouse.

In c, cpp, rust, we use as keyword to cast the numbers to ignore overflow.
In MySQL/ClickHouse, it's safe to cast a wider number to the shorter one. Users afford the overflow and accurate problem

[jorgecarleitao]

This last change was an awesome idea; it keeps it very clear (and even backward incompatible) 👍

Left some final minor comments, but the design is there.

Thanks a lot, @sundy-li , great work here!

[sundy-li]

Is it a big pain for DataFuse to not have the option?

Not yet, we just depend on the performance of cast. And I don't know whether we will have more options in cast kernel.
As SortOptions did, such like cast_null_as_default: bool, overflow_as_default: bool, so making it pub seems to be much more convinient.

[jorgecarleitao]

pulled locally and PRed minor changes. Sorry, with 2 I meant the two variations of the functions, not the 2 APIs, to keep it aligned with the other functions :)