Capture Java thread names per-sample for Spark task attribution

[prashantbytesyntax]

Summary

Enables accurate per-thread attribution in Java flamegraphs for workloads that rename threads at runtime — most importantly Spark executors, which rename their task-pool threads per task (e.g. Executor task launch worker for task 21.0 in stage 492.0).

Two changes:

• scripts/async_profiler_build_shared.sh — build async-profiler from the v3.0g1_threadpool branch (Granulate/async-profiler#10, "capture thread names for every sample"). Stock v3.0g1 records a thread's name only once, so long-lived Spark executor pool threads keep a stale task/stage label across their lifetime; the threadpool build re-reads the name at each sample.
• gprofiler/profilers/java.py — add threads to async-profiler's collapsed-output FORMAT_PARAMS so each sample is prefixed with its (now current) thread name.

Why

The Spark-on-k8s per-thread profiling demo (gprofiler-performance-studio#85) relies on thread names to break flamegraphs down by Spark stage/task. With the stock build, attribution silently drifts as executors reuse pool threads for new tasks. This PR makes the thread-name labels trustworthy.

Notes for reviewers

• This makes per-thread output always on for Java profiling (larger collapsed output / higher frame cardinality). If we'd rather keep it opt-in, the FORMAT_PARAMS hunk can be dropped and threads passed via the existing --java-async-profiler-args=threads flag instead — the build-script change alone is what fixes stale names.
• The upstream async-profiler branch is an open POC PR on Granulate/async-profiler (not yet merged); we pin an exact GIT_REV so builds are reproducible.
• Branch history contains a couple of revert/reapply cycles; the net diff is just the two files above.

🤖 Generated with Claude Code

[artursarlo]

LGTM. Tested those changes locally on a DEV environment. Working as expected.