[llvm][mustache] Use single pass when tokenizing (#159196)
The old implementation used many string searches over the same portions
of the strings. This version sacrifices some API niceness for perf wins.
| Metric | Baseline | Single-Pass | Change |
| --- | --- | --- | --- |
| Time (ms) | 36\.09 | 35\.78 | \-0.86% |
| Cycles | 35\.3M | 35\.0M | \-0.79% |
| Instructions | 86\.7M | 85\.8M | \-1.03% |
| Branch Misses | 116K | 114K | \-1.91% |
| Cache Misses | 244K | 232K | \-4.98% |
