Update primitive benches after open refactoring issues are resolved

[Panadestein]

Context

Several refactoring issues are currently open and will change core APIs, dependencies, and performance characteristics of src_method. The primitive benches (benches/primitives/) should be revisited once these land, to keep them representative and avoid measuring stale code paths.

Related open issues

• Stop configuring global logging at import; replace structlog with stdlib logging #11 Stop configuring global logging at import; replace structlog with stdlib logging
• Remove unused runtime dependencies (numba, llvmlite, cmaes) and demote cyclopts #12 Remove unused runtime dependencies (numba, llvmlite, cmaes) and demote cyclopts
• Single-precision inputs are silently promoted to double precision #13 Single-precision inputs are silently promoted to double precision
• GPU path round-trips to host on every call and every site #14 GPU path round-trips to host on every call and every site
• Validate public API arguments: chi_out=0 silently returns zero-size bonds #15 Validate public API arguments: chi_out=0 silently returns zero-size bonds
• Enforce type checking in CI: py.typed is shipped but never verified #16 Enforce type checking in CI: py.typed is shipped but never verified
• Deduplicate the four near-identical SRC sweep implementations #17 Deduplicate the four near-identical SRC sweep implementations
• Relax over-tight dependency floors and single-source the ruff pin #18 Relax over-tight dependency floors and single-source the ruff pin
• Docs: fallback behaviour is documented incorrectly and there are no examples #19 Docs: fallback behaviour is documented incorrectly and there are no examples

Task

Once the above issues are closed, review and update benches/primitives/ to:

• reflect any API/behavior changes (e.g. precision handling, argument validation, GPU execution path)
• account for dependency/logging changes affecting bench setup
• add coverage for the deduplicated SRC sweep implementation, if applicable