PyDict_GetItemString() is heavily used, especially from keyword argument parsing.
Current implementation creates temporary string for key object.
This patch avoid the temporary key string when passed C string is ASCII.

This benchmark is based on a8563ef0eb8a, so PyDict_GetItemString() calls for
parsing positional arguments is reduced already.


$ ../python -m perf compare_to -G --min-speed 2 default.json patched.json
Slower (1):
- scimark_lu: 430 ms +- 21 ms -> 446 ms +- 23 ms: 1.04x slower (+4%)

Faster (11):
- telco: 24.2 ms +- 0.4 ms -> 21.8 ms +- 0.7 ms: 1.11x faster (-10%)
- xml_etree_parse: 315 ms +- 17 ms -> 302 ms +- 14 ms: 1.04x faster (-4%)
- logging_simple: 31.6 us +- 0.3 us -> 30.4 us +- 0.3 us: 1.04x faster (-4%)
- mako: 41.6 ms +- 0.7 ms -> 40.3 ms +- 0.4 ms: 1.03x faster (-3%)
- logging_format: 36.5 us +- 0.3 us -> 35.5 us +- 0.4 us: 1.03x faster (-3%)
- float: 297 ms +- 4 ms -> 289 ms +- 4 ms: 1.03x faster (-3%)
- scimark_monte_carlo: 276 ms +- 10 ms -> 269 ms +- 7 ms: 1.02x faster (-2%)
- regex_effbot: 5.31 ms +- 0.37 ms -> 5.19 ms +- 0.06 ms: 1.02x faster (-2%)
- pickle_pure_python: 1.32 ms +- 0.02 ms -> 1.29 ms +- 0.02 ms: 1.02x faster (-2%)
- scimark_sor: 525 ms +- 9 ms -> 514 ms +- 8 ms: 1.02x faster (-2%)
- richards: 180 ms +- 3 ms -> 176 ms +- 2 ms: 1.02x faster (-2%)

Benchmark hidden because not significant (52): ...


Performance difference of telco is bit surprising.
Profiler shows the difference is from `print(t, file=outfil)` (here: https://github.com/python/performance/blob/master/performance/benchmarks/bm_telco.py#L79 )

Until most common builtin functions are converted to FASTCALL, this patch has significant
performance gain.