As mentioned in #276, current behavior is that short strings can be stored as static strings or inline. This adds an overhead for interning short strings, as they need to be hashed and compared to possible matches in the static set.

This PR changes the behavior to always store strings of 7 bytes or less inline.

This greatly improves the performance of interning short strings - approx 250% speedup, depending on length of the string.

The benchmarks in this repo show no apparent negative effect on any other operations (though there's a lot of noise in the benchmarks).

This PR does not involve any breaking changes.

Even though interning short static atoms is very fast anyway, I imagine this change will translate to a significant real-world performance improvement for some use cases. For example, SWC uses string-cache when parsing Javascript which typically includes a lot of short variable names and tokens, especially in minified code e.g. for (let i = 0; i < len; i++).

Notes on implementation

• The empty string is still stored in static set. This is slightly non-optimal, but avoids a breaking change (StaticAtomSet::empty_string_index would need to be removed otherwise).
• Strings <= 7 bytes can still be passed to the codegen but it generates pack_inline() calls in place of pack_static() to generate inline atoms at compile time.
• Benchmarks needed to be altered to use longer strings when benchmarking static atoms. Benchmark changes are in a separate commit to allow comparing the benchmarks before and after the substantive change.
• This PR also includes the changes in Faster deref of inline atoms #277. Without this, this PR produced a performance degradation deref-ing short strings as inline was not as performant as static.