Make get_short_name return Cow instead of String

[nicopap]

Objective

• Do not make allocation in get_short_name when not necessary. The get_short_name of a path without generics is strictly a subslice of the path itself, so we don't need to create a new buffer for it.
• Improve drastically parsing performance by working on ASCII (this allows skipping UTF8 checks (linear with string size) on every slice splits) It is still fully UTF8-complient, for the same reasons as Make the reflect path parser utf-8-unaware #9371
• Fixes get_short_name should return a Cow most likely disqualified#2

Solution

• Rework get_short_name to work on &[u8] instead of &str, only returns an owned String when the path contains a special character.
• Added DisplayShortName. It stores any kind of string, but when displayed, it displays the get_short_name version.

Future work

I think we could realistically replace the GenericTypePathCell with a streaming version of get_short_name that implements Hash, Eq etc. similarly to DisplayShortName. It would add the parsing cost on top of hashing/equality check, but now that the parser is a lot faster, it's questionable whether the cost is actually significant.

We could also replace the CowStr in the scheduling error type with DisplayShortName, to lazily parse the type path only when displaying it.

---

Changelog

• get_short_name returns a Cow<str> instead of String, avoiding allocation when not strictly necessary. It is also significantly more performant
• Added DisplayShortName. It stores any kind of string, but when displayed, it displays the get_short_name version.

Migration Guide

• get_short_name returns a Cow<str> instead of String, add .to_string() to the return value to get a String. However, we recommend that if you can, you store the Cow over the String to avoid allocations.

[nicopap]

On the performance-side of things. I'll assume it's more perfs because:

1. It's the same optimization as Make the reflect path parser utf-8-unaware #9371, and it has been proven to be a 20% speedup on benchmarks
2. In the simple case, we completely skip allocation
3. We pre-allocate to a size known to be able to hold the return value. For a 33 chars string, that saves us 2 reallocations, for 65 chars: 3 etc.
4. We might elide bound checks (though they were probably mitigated by the optimizer already) by storing the &[u8] instead of storing an offset and getting the sub-slice.

[alice-i-cecile]

I like this change, and agree that it should improve performance without seriously hurting complexity. Can you add a test for a non-ASCII type name to verify that it works correctly (and continues to in the future)?

[hymm]

Could you rebase or merge main? There's some changes I made to schedule.rs that are conflicting. sorry 😅

[nicopap]

There is also the uppercase fix (#9587), which I'm not exactly sure how to integrate in this change.

[alice-i-cecile]

@nicopap want me to toss Adopt-Me on this?

[nicopap]

@nicopap want me to toss Adopt-Me on this?

Sure. Should be easy to pick up.

[james7132]

Closing in favor of #12275.