Optimize dependency iteration

[killerdevildog]

…cies

PROBLEM IDENTIFIED:
Before this optimization, dependencies were being parsed twice during candidate resolution:

1. During _check_metadata_consistency() for validation (line 233 in original code)
2. During iter_dependencies() for actual dependency resolution (line 258)

This caused significant performance issues because:

• dist.iter_provided_extras() was called multiple times
• dist.iter_dependencies() was called multiple times
• Parsing requirements from package metadata is computationally expensive
• The TODO comment at line 230 specifically noted this performance problem

SOLUTION IMPLEMENTED:
Added caching mechanism with two new instance variables:

• _cached_dependencies: stores list[Requirement] after parsing once
• _cached_extras: stores list[NormalizedName] after parsing once

HOW THE CACHING WORKS:

1. Cache variables are initialized as None in init()
2. During _prepare() -> _check_metadata_consistency(), dependencies are parsed and cached during validation
3. During iter_dependencies(), the cached results are reused via _get_cached_dependencies()
4. Cache is populated lazily - only when first accessed
5. Subsequent calls to iter_dependencies() use cached data (no re-parsing)
6. Each candidate instance has its own cache (thread-safe)

ADDITIONAL OPTIMIZATIONS:

• Also optimized ExtrasCandidate.iter_dependencies() to cache iter_provided_extras() results
• Ensures consistency between validation and dependency resolution phases

TESTING PERFORMED:

1. Created comprehensive test script (test_performance_optimization.py)
2. Used mock objects to verify iter_provided_extras() and iter_dependencies() are called at most once
3. Verified pip install --dry-run works correctly with caching
4. Test results showed 0 additional calls to parsing methods during multiple iter_dependencies() invocations
5. Functional testing confirmed dependency resolution still works correctly

PERFORMANCE IMPACT:

• Eliminates duplicate parsing during metadata consistency checks
• Reduces CPU time for packages with complex dependency trees
• Especially beneficial for packages with many dependencies
• Memory overhead is minimal (only stores parsed results, not raw metadata)

Resolves TODO comment about performance in candidates.py line 230

[notatallshaw]

Thanks for your PR to pip. Please be aware all maintainers are volunteers so it may take a moment for someone to review.

Let me know if you need any help fixing the the linting and pre-commit errors, you should be able to locally run them following this guide: https://pip.pypa.io/en/stable/development/getting-started/#running-linters

Also, are there real world scenarios that inspired you to fix this? Or is this more of an academic exercise?