Make StringDictionaryBuilder faster

[tustvold]

Is your feature request related to a problem or challenge? Please describe what you are trying to do.

A while back I implemented an optimized string dictionary builder for IOx. This contains two major tricks to provide better performance:

• Use ahash instead of SipHash - this alone provides a 40% speedup
• Use hashbrown's raw_entry_mut to not duplicate string values into the hashmap

I have an implementation of this for arrow that needs a bit more polish, but leads to a 60% speedup over the current implementation in arrow. Unfortunately it depends on #1850 as it needs to be able to read the string data from an in-progress StringBuilder

Describe the solution you'd like

Implement #1850 and then add this functionality

Describe alternatives you've considered

We could not do this

[jhorstmann]

I also have an implementation that I could contribute, it would need some changes for generic key types and maybe for null values. It's based on using hashbrown directly and storing a (start_index, end_index) tuple in the hashmap into a backing MutableBuffer, also keeping track of the offsets in a MutableBuffer. Probably very similar to your impl, but using mutable buffers directly.

[tustvold]

Exciting, lets see how they compare 😄