bench: add micro-benchmarks for fastcall and tp_vectorcall paths

Replaces large-buffer-only benchmarks with a mix that captures
the recent optimizations:
- 5-byte inputs (call overhead dominates, 25-60% improvement)
- str input (tests _get_buffer_or_str UTF-8 path)
- type constructors (tests tp_vectorcall, 41-66% improvement)
- keyword seed passing
- Retains 1KB/10KB macro benchmarks for covering hash throughput

Author: ifduyue