Support plain mixed-number extraction

[dkajtoch]

Problem

math_verify.parse() does not recognize plain mixed-number notation such as 11 5/6 as one value. The generic expression and number patterns can instead expose finite fragments of the answer as independent candidates.

This causes more than a missed equivalence with 71/6: it can produce false positives. For example, 2 2/3 can expose 2 and incorrectly match a gold answer of 2. Invalid mixed-looking syntax has the same problem: 4 10/3 can expose 10 and incorrectly match a gold answer of 10 instead of being rejected.

Solution

This PR adds a higher-priority mixed-number extraction path before the generic expression and number patterns.

• Recognize an optional sign, an integer whole-number component, horizontal whitespace, and a fractional component.
• Require the fractional component to be proper: the denominator must be positive and the numerator must be smaller than the denominator.
• Normalize valid values to an explicit sum before parsing, with a leading negative sign applying to the complete value.
• Preserve the original matched text as the fallback representation.
• Return a non-finite result for invalid mixed-looking forms so extraction cannot fall through to a finite prefix or numerator.
• Apply percentage suffixes to the complete mixed-number value.
• Do not infer an operation when the left component is a fraction, decimal, parenthesized expression, or appears on another line. Existing explicit addition and subtraction remain unchanged.

Consequently, 11 5/6 is interpreted as 11 + 5/6, while 4 10/3 is rejected. The explicit expression 4 + 10/3 remains valid.

Checks run

• .venv/bin/pytest -q tests/test_mixed_number_extraction.py: 42 passed
• .venv/bin/pytest -q: 332 passed, 1 failed; the remaining LaTeX-percentage failure is unchanged on upstream/main
• .venv/bin/ruff check tests/test_mixed_number_extraction.py: passed
• .venv/bin/ruff format --check src/math_verify/parser.py tests/test_mixed_number_extraction.py: passed
• uv build: source distribution and wheel built successfully

The existing parser.py import-order and unused-import Ruff findings are also present on upstream/main and are outside this change.

[dkajtoch]

@hynky1999, would you mind reviewing this change? It adds plain mixed-number extraction and prevents finite-fragment false positives such as 2 2/3 matching gold 2. The focused suite has 42 passing tests; the only full-suite failure is the pre-existing LaTeX-percentage case documented in the PR description.