perf: Limit tokens per pattern to 128

[benclive]

What this PR does / why we need it:

• Limits the number of tokens generated per log-line for use in pattern detection
• Some log lines are extremely long, which generate very specific patterns & use a lot of memory to store.
• The patterns themselves aren't useful as they are so specific, so here the number of tokens is limited to 128.
  • Based on testing, this is ~p95 of patterns so the majority of patterns are unaffected.
• Also took the opportunity to use append instead of direct assignment, as it's easier to read.

With this change, the pattern-ingester component OOMs significantly less than before: