Words in all capitals use too much storage in Elasticsearch

I noticed that words like ALL_CAPITALS_WORD are giving the following tokens:

"all_capitals_word"
"all"
"ll"
"capitals"
"apitals"
"pitals"
"itals"
"tals"
"als"
"ls"
"word"
"ord"
"rd"

In #27918 (closed) we agreed that partial words don't need to be searchable and the (?=([\\p{Lu}]+[\\p{L}]+)) pattern happens to be creating all the partial word tokens (starting from the end) for capital words.

The ideal tokens for this word should be:

"all_capitals_word"
"all"
"capitals"
"word"

We could likely get some large storage savings by improving that pattern. I believe that pattern is also responsible for splitting up words like AllCapitalsWord which is why it is splitting on every capital letter. It may be difficult to distinguish those cases.