feat(node,cli): expose on-demand pprof HTTP endpoint

[ByteYue]

Summary

The node previously supported only a hard-coded periodic CPU profiler (ENABLE_PPROF=1) that writes .pb files to disk every three minutes. That works for post-mortems but is awkward when you want to profile a specific interval — you have to scp the file off the node and match it to wall-clock by hand.

This PR adds an on-demand HTTP endpoint and a CLI wrapper.

Node side

• New --pprof_addr <ADDR> flag on GravityNodeArgs (e.g. 127.0.0.1:6060). When set, starts an axum HTTP server on the existing tokio runtime exposing:
  • GET / — index page documenting the endpoints
  • GET /debug/pprof/profile?seconds=N[&frequency=Hz] — returns a protobuf CPU profile consumable by go tool pprof
• Concurrency: a process-wide mutex serializes overlapping requests. pprof uses the global SIGPROF handler; overlapping ProfilerGuards produce garbage data, so the second request gets 409 Conflict.
• Interaction with existing mode: --pprof_addr disables the periodic disk-dump mode (ENABLE_PPROF=1) because they conflict over the same profiler state. A warning is emitted if both are set.

CLI side

gravity-cli node pprof cpu [--addr ADDR] [--duration SECS] [--frequency HZ] [--output-file PATH]

• --addr defaults to 127.0.0.1:6060 (matches common local setup), overridable via GRAVITY_PPROF_ADDR.
• --output-file - streams protobuf bytes to stdout.
• On success prints the go tool pprof -http=:8080 <file> command to view the flamegraph.

Scope

• Heap profiling is out of scope. Exposing jemalloc heap snapshots requires coordinating MALLOC_CONF prof_prefix with the HTTP server and reading back dump files; deferred to a follow-up.
• No changes to the periodic ENABLE_PPROF mode except the new precedence rule.

Test plan

• Integration test pprof_server_end_to_end (in-process): binds ephemeral port → index returns 200 + documents endpoint → profile returns 200 with application/octet-stream and valid protobuf tag byte → concurrent overlapping request returns 409 while the first succeeds with 200
• gravity_node --help lists --pprof_addr with full description
• gravity-cli node pprof cpu --help lists all flags
• CLI error path: connecting to non-listening port → clean error with hint, exit 1
• cargo build -p gravity_node -p gravity_cli --profile quick-release succeeds with RUSTFLAGS="--cfg tokio_unstable"