test(benchmarks): add macro-heavy Jinja baseline

[eitsupi]

Summary

• add a deterministic macro-heavy cache benchmark profile with 64 models, 16 macro source files, and 128 valid macros
• preserve the existing small and medium workload shapes
• validate macro source/definition counts and expose them in benchmark summaries
• run semantic cache validation in CI for both small and macro-heavy profiles without timing gates
• document the profile for measuring cumulative macro validation and per-model prefix parsing

Baseline

Release binary built from 3723b79, hyperfine 10 runs / 2 warmups:

SQL profile	No cache	Persistent cold	Persistent warm
small (64 models, 1 file, 1 macro)	9.972 ms	11.094 ms	9.669 ms
medium (512 models, 1 file, 1 macro)	18.813 ms	20.076 ms	16.819 ms
macro-heavy (64 models, 16 files, 128 macros)	32.670 ms	47.488 ms	12.167 ms

All SQL, manifest, and column semantic probes passed for every profile. SQL file, macro, and vars invalidation probes also passed.

Validation

• deterministic generation self-check for small, medium, and macro-heavy
• workload validation for all three profiles
• timed cache benchmark for all three profiles
• local CI-equivalent --skip-timing validation for small and macro-heavy
• Python bytecode compilation
• git diff --check