All tables use the compact format now, and we generate
caches in that format too. Bump the cache version to 3
for this.
Replace the python script for generating the builtin table
by a small C program using the same code to generate the data
for the builtin table. This drops the restriction on only
generating a single character in the builtin sequences.
This lets us naturally replace matching sequences
while parsing. That means that the semantics are now
"last one wins" if the parser sees multiple entries
for the same sequence.
Add a testcase that checks the new replacement semantics.
