Optimise SmolStr::clone 4-5x speedup inline, 0.5x heap (slow down)

[alexheretic]

This is an interesting performance trade-off. We can significantly improve inline SmolStr::clone performance at the expense of heap performance if we use #[cold] & #[inline(never)] hints against the latter.

This is a better result for users that care most about inline variant performance. If we think that represents most users it may be worth it.

Benchmarks

• inline clones 4.1ns -> 0.8ns 🙂
• heap clones 2.8ns -> 5.8ns 🙁

group                      master                  pr
-----                      ------                  --
SmolStr::clone len=12      4.86      4.1±0.00ns    1.00      0.8±0.00ns
SmolStr::clone len=50      1.00      2.8±0.02ns    2.05      5.8±0.01ns
SmolStr::clone len=1000    1.00      2.8±0.01ns    2.04      5.8±0.02ns