[reese]

This is another change to try and shrink the token target types, in this case ConditionalLayoutEntry. ConditionalLayoutEntry currently holds two vectors, a pointer to the keyword slice, an indent depth (u32), and the phase enum. This PR combines the two vectors into one (and storing an index where the two diverge) and then storing the keyword as an enum, such as we did for delimiters in #981. This gets it down to 40 bytes each, which gets our overall memory usage down another ~15%.

dhat: Total:     15,629,666 bytes in 44,105 blocks
dhat: At t-gmax: 516,575 bytes in 859 blocks

[froydnj]

I kind of wish there was a better way of consuming self here so that the abstraction of predicate and statement tokens would not have to leak into write_inline_tokens and write_block_tokens. Like the consumption of self above would actually return two iterator-ish things directly, rather than manually splitting here.