feat: extend gcprofile with --by=class and --fold (allocation hotspot views)

[rlaope]

Closes #152.

Summary

Extends argus gcprofile with two new views instead of introducing a redundant argus alloc command. During implementation it became clear that gcprofile already does 80% of what the original alloc proposal described — same JFR event types (jdk.ObjectAllocationInNewTLAB / OutsideTLAB), same --duration / --top, same top-N table, same JSON. A second command would just split the mental model.

What gcprofile was actually missing (vs async-profiler -e alloc / JDK Mission Control):

1. A class-grouped view — today it only shows stack-frame sites.
2. Flamegraph-compatible output — today there is no way to visualize call chains.

This PR closes both gaps with two surgical flags.

Changes

• --by=site|class (default site — unchanged)
  • site — current stack-frame aggregation.
  • class — aggregates by the allocated object's type (event.getClass(\"objectClass\")). Answers "which classes are burning bytes".
• --fold=FILE — writes folded stacks in Brendan Gregg's flamegraph.pl format (root-first, leaf-last, merged by total bytes). Render with flamegraph.pl --colors=mem FILE > alloc.svg.
• New public API in AllocationProfiler:
  • analyzeByClass(Path) → AllocationByClass
  • analyzeFoldedStacks(Path) → Map<String, Long>
• Help text and bash completion updated.

Non-goals in this PR

• Not a new command. argus gcprofile remains the single entry point for allocation profiling.
• No inline SVG rendering — --fold=FILE produces folded text; SVG is one flamegraph.pl call away. A bundled renderer can be added in a follow-up if needed.
• No live streaming — bounded --duration, same as today.

Backwards compatibility

argus gcprofile <pid> (no extra flags) emits the exact same output as before. No existing caller is affected.

Developer experience

# Unchanged — default stack-frame view
argus gcprofile 12345 --duration=30s

# New — allocated types
argus gcprofile 12345 --duration=30s --by=class
#   1. byte[]                       1.8 GB  42.1%
#   2. java.util.HashMap\$Node       620 MB  14.4%

# New — folded stacks for flamegraph.pl
argus gcprofile 12345 --duration=30s --fold=alloc.folded
flamegraph.pl --colors=mem alloc.folded > alloc.svg

References

• async-profiler -e alloc mode
• JFR: jdk.ObjectAllocationInNewTLAB, jdk.ObjectAllocationOutsideTLAB
• Brendan Gregg, "Folded stacks" format (flamegraph.pl)

Test plan

• ./gradlew :argus-cli:compileJava — BUILD SUCCESSFUL
• ./gradlew :argus-cli:test — full suite passes
• New unit tests cover: byClass sort/rate/coalesce/empty; folded stacks reordering/merging/sorting/skip-empty
• Existing argus gcprofile behavior unchanged (default --by=site path still the first branch)