Reuse the latest-version parse instead of parsing twice

[Janther]

Slang's inferLanguageVersions internally parses the whole file using the latest parser they have and fetches all VersionExpressionSets to then return an array of matching versions.

The way we were parsing, we were doing the work twice, once for inference and once for the actual parse.
By parsing directly with the latest parser and collecting the pragmas, we can use a much smaller input to infer the same array of versions and then if we happen to match with the latest version, we have an output ready to go.

As an example for OpenZeppelin v3.4's TimelockController.sol (301 lines), the inference receives only:

// SPDX-License-Identifier: MIT

pragma solidity >=0.6.9 <0.8.0;

pragma experimental ABIEncoderV2;

The highest version these pragmas allow is 0.7.6, not the latest, so we parse again with 0.7.6. That's the same two parses as before, plus collecting the pragmas through calls from JavaScript into Slang instead of entirely inside Slang. That overhead is small enough to be lost in the noise here (15.2 ms → 15.1 ms).

Pragmas anywhere in the file are found, which matters for flattened files. For OpenZeppelin's ERC20 flattened with its dependencies (604 lines, 7 contracts), it receives only:

// File: …/token/ERC20/IERC20.sol
// SPDX-License-Identifier: MIT
pragma solidity >=0.4.16;

// File: …/token/ERC20/extensions/IERC20Metadata.sol
pragma solidity >=0.6.2;

// File: …/utils/Context.sol
pragma solidity ^0.8.20;

// File: …/interfaces/draft-IERC6093.sol
pragma solidity >=0.8.4;

// File: …/token/ERC20/ERC20.sol
pragma solidity ^0.8.20;

Here the highest allowed version is the latest one (0.8.37), so the parse we already have is the result: 1 parse instead of an inference plus a parse (12.5 ms → 6.7 ms).

In the worst case, the inferred array will force us to parse again with an older version which is roughly the same amount of work we were originally doing.

File	Work	Compared with main
The pragmas allow the latest version	1 parse	~30% faster
The pragmas need an older version (e.g. 0.8.26), or the latest version can't parse the file (syntax from before 0.8.0)	2 parses, plus collecting the pragmas from JavaScript	about the same

createParser also now returns { version, parseOutput } instead of the Parser, since callers only ever read parser.languageVersion.

The name createParser is not accurate anymore, since we return { version, parseOutput }, so I would be changing it in a PR after this since it touches a few files in testing as well. I'm thinking about parseSourceUnit() since that describes better what it does. Welcoming suggestions @fvictorio

Benchmark

Formatting OpenZeppelin v5.7.0's contracts/ (368 files) with Prettier 3.9.9 on Node 26 (Apple M3 Pro):

hyperfine --warmup 2 --runs 10 --ignore-failure \
  -n main "npx prettier --check --no-config --plugin=./main/dist/index.js 'contracts/**/*.sol'" \
  -n pr   "npx prettier --check --no-config --plugin=./pr/dist/index.js 'contracts/**/*.sol'"

	Time
main	3.110 s ± 0.020 s
This PR	2.381 s ± 0.023 s

1.31× faster.