Avoid per-token source index visitor calls (#26506)
## Summary
`Indexer::from_tokens` previously called both range builders and checked
string flags for every token, even though only string tokens can affect
those indexes or span multiple lines.
Dispatch the existing builders from the token-kind match instead.
Interpolated-string boundaries visit the interpolated-range builder,
regular strings and f-string middles visit the multiline-range builder,
and t-string middles only update the current line start. This preserves
the indexed ranges while removing visitor calls and flag lookup from the
common non-string path.
The empty-trivia fast path was split into #26507 so CodSpeed can measure
the two optimizations independently.
## Performance
The [full CodSpeed
report](https://app.codspeed.io/astral-sh/ruff/branches/charlie%2Fcodex-parser-throughput)
compares `f673f29` with `dc42d1d`. All 15 linter CPU-simulation
measurements improve:
| Configuration | Gain across the five workloads |
| --- | ---: |
| Default rules | 1.52-4.63% |
| All rules | 0.36-1.09% |
| All rules with preview | 0.31-0.89% |
Summing the five default-rules head and base measurements gives a 4.00%
aggregate improvement. CodSpeed classifies
`linter/default-rules[large/dataset.py]` at 4.63% as a significant
improvement; the other measured gains are below its reporting threshold.
