perf(trace): optimize traceSingleQuotedString

[analytically]

Replace fmt.Fprintf with direct hex encoding using lookup table. Add Grow() pre-allocation to eliminate builder reallocations.

Benchmarks show 1.7-8x speedup depending on input:

• all non-printable: 743ns -> 92ns (8x)
• mixed content: 422ns -> 108ns (4x)
• printable only: 242ns -> 144ns (1.7x)

Allocations reduced from 2-6 to 1 in all cases.

[jackc]

Take a look at #2457. I'm still reluctant to merge optimizations of paths that shouldn't be triggered very often, but if it is to be done then I think that #2457 is a superset of this.