Refactor: compact and optimize infer_overload_return_type while preserving behavior and comments

[victorletichevsky]

This pull request contributes to issue #5917, which encourages the decomposition and simplification of overly long methods to improve code clarity and maintainability.

Summary of changes

Refactored the infer_overload_return_type method to reduce line count and eliminate structural redundancy, while preserving all original logic and explanatory comments.

Key improvements:

• Reduces the overall number of lines without removing any original inline documentation
• Preserves all key comments explaining overload matching logic, especially around Any-based ambiguity
• Replaces nested ifs with continue for early exits, minimizing indentation
• Groups related variable declarations to improve readability
• Avoids unnecessary allocations when possible (e.g. skipping data collection if early return applies)

Notes

• This is a pure refactor in line with issue Find longest functions and split them #5917.
• The logic has not been modified, only cleaned up for readability and maintainability.

[github-actions]

Diff from mypy_primer, showing the effect of this PR on open source code:

typeshed-stats (https://github.com/AlexWaygood/typeshed-stats): 1.41x slower (65.1s -> 91.6s in single noisy sample)

nox (https://github.com/wntrblm/nox): 1.22x slower (23.7s -> 28.9s in single noisy sample)

materialize (https://github.com/MaterializeInc/materialize): 1.08x slower (129.8s -> 140.7s in single noisy sample)

discord.py (https://github.com/Rapptz/discord.py): 1.07x slower (275.7s -> 296.1s in single noisy sample)

[sterliakov]

Why new cast?