gh-161469: Add re benchmarks and _sre tail-call jump dispatch - #36

Summary

• add a new Tools/scripts/re_bench.py suite covering multiple re features (literal, charset, quantifier, alternation, backref, lookaround) across multiple input sizes
• document the new benchmark script in Tools/scripts/README
• add an initial _Py_TAIL_CALL_INTERP path in Modules/_sre/sre_lib.h by applying tail-call dispatch to DO_JUMP transitions, while preserving the existing non-tail-call and computed-goto behavior
• add ceval-style compiler gating for musttail + preserve_none in the new path

Validation

• built with tail-call interpreter enabled:
  • CC=clang-19 ./configure --with-tail-call-interp --without-ensurepip
  • make -j4
• verified compiler tail-call attribute support:
  • clang-19 -O3 -Werror -c /tmp/tailcall_probe.c
• verified backwards compatibility build path:
  • CC=gcc /workspaces/cpython/sobolevn/cpython/configure --without-ensurepip
  • make -j4 Modules/_sre/sre.o
• ran tests:
  • ./python -m test -j2 test_re

Benchmark data

Command used for before/after:

• ./python Tools/scripts/re_bench.py --seconds 0.4 --repeat 5 --write /tmp/re-baseline.json
• ./python Tools/scripts/re_bench.py --seconds 0.4 --repeat 5 --read /tmp/re-baseline.json --write /tmp/re-after.json

Notable deltas (best-of-5):

• alternation-search: +4.6% to +9.1% across sizes
• most other cases: between ~0% and moderate regressions in this first prototype run
• geometric mean over all 24 case/size entries: 0.983x (about -1.7%)

This is an initial experiment to establish reproducible measurement and a minimal tail-call integration point for _sre; follow-up tuning can iterate on the dispatch strategy.