This optimization is only for short strings.  There is no significant difference for long and non-ASCII strings.

```
# 1000 ASCII
$ ./python -m pyperf timeit --compare-to=./python-master -s 'b=b"f"*1000' -- 'b.decode()'
python-master: ..................... 196 ns +- 1 ns
python: ..................... 185 ns +- 1 ns

Mean +- std dev: [python-master] 196 ns +- 1 ns -> [python] 185 ns +- 1 ns: 1.06x faster (-6%)

# 10000 ASCII
$ ./python -m pyperf timeit --compare-to=./python-master -s 'b=b"f"*10000' -- 'b.decode()'
python-master: ..................... 671 ns +- 4 ns
python: ..................... 662 ns +- 1 ns

Mean +- std dev: [python-master] 671 ns +- 4 ns -> [python] 662 ns +- 1 ns: 1.01x faster (-1%)

# 100000 ASCII
$ ./python -m pyperf timeit --compare-to=./python-master -s 'b=b"f"*100000' -- 'b.decode()'
python-master: ..................... 5.86 us +- 0.03 us
python: ..................... 5.85 us +- 0.02 us

Mean +- std dev: [python-master] 5.86 us +- 0.03 us -> [python] 5.85 us +- 0.02 us: 1.00x faster (-0%)

# 1 non-ASCII
$ ./python -m pyperf timeit --compare-to=./python-master -s 'b="あ".encode()' -- 'b.decode()'
python-master: ..................... 138 ns +- 2 ns
python: ..................... 136 ns +- 2 ns

Mean +- std dev: [python-master] 138 ns +- 2 ns -> [python] 136 ns +- 2 ns: 1.02x faster (-2%)

# 1000 ASCII + 1 non-ASCII
$ ./python -m pyperf timeit --compare-to=./python-master -s 'b=b"x"*1000 + "あ".encode()' -- 'b.decode()'
python-master: ..................... 361 ns +- 9 ns
python: ..................... 360 ns +- 5 ns

Mean +- std dev: [python-master] 361 ns +- 9 ns -> [python] 360 ns +- 5 ns: 1.00x faster (-0%)
Not significant!

# 10000 ASCII + 1 non-ASCII
$ ./python -m pyperf timeit --compare-to=./python-master -s 'b=b"x"*10000 + "あ".encode()' -- 'b.decode()'
python-master: ..................... 2.83 us +- 0.02 us
python: ..................... 2.83 us +- 0.03 us

Mean +- std dev: [python-master] 2.83 us +- 0.02 us -> [python] 2.83 us +- 0.03 us: 1.00x slower (+0%)
Not significant!
```