tracing: keep stacks compressed in tracer snapshots

We have seen some production instances with nearly 100k goroutines where the stack dump alone is >64MiB. Currently a tracer snapshot stores a map of every goroutine to its raw string stack; even just a few snapshots thus could add up to non-trivial memory usage. Instead, we could store the dump bytes compressed and only decompress them to parse out individual goroutine IDs and their stacks when needed.

Jira issue: CRDB-25099