Improve flamegraph profiling: deeper stacks, off-CPU, better CSV

[vadv]

Summary

• Use dwarf,16384 stack dump size for complete tokio async call stacks (default 8192 truncates deep stacks)
• Add FLAMEGRAPH_OFFCPU=1 mode — generates blue flamegraph showing where pg_doorman waits (locks, pool contention, I/O) via offcputime-bpfcc
• Generate profile.csv with both self and inclusive cost per function, plus summary.csv with CPU breakdown by category (pg_doorman, tokio, allocator, kernel_tcp, kernel_nftables, etc.)

Usage

# Standard CPU flamegraph
make flamegraph

# With off-CPU analysis
FLAMEGRAPH_OFFCPU=1 make flamegraph

New artifacts

flamegraph-output/<timestamp>/
├── profile.csv              # function,self_samples,self_pct,inclusive_samples,inclusive_pct,category
├── summary.csv              # category,self_samples,self_pct
├── flamegraph-offcpu.svg    # (only with FLAMEGRAPH_OFFCPU=1)
└── ...existing artifacts...

Test plan

• make flamegraph produces all artifacts including new CSVs
• profile.csv has correct inclusive cost (entry_SYSCALL at 72%, pg_doorman::handle at 63%)
• summary.csv shows meaningful category breakdown
• Cleanup works, no zombie processes
• FLAMEGRAPH_OFFCPU=1 make flamegraph with bcc-tools installed