Signed-off-by: Brooks Townsend <[email protected]>

This failed since 464b1f7:

  (cd idna && cargo test --all-targets --no-default-features --features alloc -- --ignored)

* Add AsciiSet::EMPTY and impl ops::Add for AsciiSet

In RFCs, the sets of characters to percent-encode are often defined as
the union of multiple sets. This change adds an `EMPTY` constant to
`AsciiSet` and implements the `Add` trait for `AsciiSet` so that sets
can be combined with the `+` operator.

AsciiSet now derives `Debug`, `PartialEq`, and `Eq` so that it can be
used in tests.

* implement ops::Not for AsciiSet

* Add const functions for negation / union of AsciiSet

* refactor: move AsciiSet related code to ascii_set.rs

Makes it a little easier to just look at the code related to the set rather than encoding

* Add tests for percent_encoding

* Add a test for issue 974

* Fix the issue

* rename test

Signed-off-by: Brooks Townsend <[email protected]>

In our tests using the serde impl on `Url`, we found that failing URL serialization says the inverse of what it should say.

```toml
[project]
name = "foo"
version = "0.0.0"
dependencies = [
  "tqdm ==4.66.0",
]
[tool.uv.sources]
tqdm = { url = "§invalid#+#*Ä" }
```

```
error: Failed to parse: `pyproject.toml`
  Caused by: TOML parse error at line 10, column 16
   |
10 | tqdm = { url = "§invalid#+#*Ä" }
   |                ^^^^^^^^^^^^^^^^^
invalid value: string "§invalid#+#*Ä", expected relative URL without a base
```

It says that expected a relative URL without a base, when this was the unexpected input that caused the error. Using `serde::de::Error::custom` oder `serde::de::Error::invalid_value` improves the error message:

```
error: TOML parse error at line 8, column 16
  |
8 | tqdm = { url = "§invalid#+#*Ä" }
  |                ^^^^^^^^^^^^^^^^^
relative URL without a base: "§invalid#+#*Ä"
```

Signed-off-by: Arthur Chan <[email protected]>

* Adjust Punycode overflow checks

* The change made in 1.0.0 incorrectly assumed that the input length
  limit removed the need to do overflow check when decoding. Now the
  internal-caller length limit is taken as a permission to skip
  overflow checks only when encoding.
* The RFC gives overflow checking pre-flight math for languages like
  that don't have checked math. Since Rust does, the code now uses
  checked_add and checked_mul instead of pre-flight when overflow
  checks are performed.

* Remove no_std category (crates.io doesn't support it, and it is now rejected), use keywords instead

* Add benches that use the main idna 1.0 entry point in idna and url

* Put the Unicode back end behind an adapter crate

* Split fastest ASCII fast path from the rest

* Bench hyphen in a domain that is otherwise lower-case ASCII

* Adjust MSRV

* Add README remarks about alternative Unicode back ends

* Change the idna_adapter dependency to crates.io

* Address clippy lints

* Increment version number of idna to 1.0.3

* Test MSRV with idna unicode-rs back end and test ICU4X back end with 1.67

* Prepare url crate for publication with idna 1.0.3 (#987)

---------

Co-authored-by: Manish Goregaokar <[email protected]>
Co-authored-by: Valentin Gosu <[email protected]>