Optimize Transliteration runtime (and add better benchmarks)

[skius]

Add better benchmarking inputs to transliteration, then:

Address comments linked to this issue in the runtime code, in particular

• does the replacement size estimate help?
• does binary searching the variable table help?
• does special casing lookup_matcher/lookup_replacing in VarTable help?
• Can we have a more efficient Insertable implementation by preemptively pushing the frozen tail to the end of the buffer's capacity when reallocating?
• do we want to implement Insertable with a rope/cord?
• See ICU4J freeze() https://github.com/unicode-org/icu/blob/388b768262684300bf47bcee8796a9327ccc2652/icu4j/main/classes/translit/src/com/ibm/icu/text/TransliterationRuleSet.java#L92