perf: add cold-start CodSpeed scenarios + tight-polling readiness (refs #557)

[Tagar]

Summary

Adds 4 CodSpeed scenarios that map 1:1 to py4j cold-start improvement work areas (refs #557), plus tightens the perf-framework JVM readiness check so cold-start scenarios get enough samples per CodSpeed budget.

Pure measurement infrastructure — zero behavior change in py4j proper. Touches only the perf testing subtree.

This is the foundation for a follow-up PR adding the actual cold-start fix (attribute memoization, ~50 LOC) — landing the metrics first means the follow-up's PR description can cite "the XA-3 scenario added here now shows ×200 improvement on this branch" rather than asking reviewers to trust unmeasured claims.

New scenarios

ID	What	Why
XA-3	1 000 walks of gateway.jvm.java.lang.System	Attribute-resolution cache canary. Each walk re-issues 3 reflection round-trips today; should collapse to ~0 RTTs after first walk once the upcoming memoization PR lands.
XB	50 fresh JavaGateway() against the fixture JVM per round	Per-connection setup cost. Surfaces Java-side accept-loop allocations + Python-side _create_connection + socket setup.
XC	X1-1 workload with CallbackServer started but no callbacks invoked	Delta vs X1-1 = steady-state callback-infrastructure overhead.
XD	subprocess.Popen → first call → shutdown, one cycle per round	Full cold start. Target of any JVM-flag work (AppCDS, -XX:TieredStopAtLevel=1, CRaC).

XA/XB/XC share the long-running fixture JVM through test_macro_scenario. XD owns its full JVM lifecycle inside measure() and is dispatched via a separate test_cold_start_scenario test function — it can't share the fixture because both default to port 25333.

Readiness model overhaul

perf.jvm.fresh_jvm shifts from blind sleep to tight polling:

	Before	After
startup_sleep	0.25 s	0.0 s
readiness_retries	1	300
poll interval	2.0 s (hardcoded)	0.05 s
Total ceiling	2.25 s	15 s

Why this matters: under the old model each XD iteration cost ~2.7 s, leaving CodSpeed only 1 sample per benchmark budget — too few for variance / regression detection. Under the new model XD per-iteration finishes in ~500-1100 ms (host-dependent), enough for 2-5 samples per budget.

The unconditional 0.25 s startup_sleep's original rationale ("let the OS reuse the listen port") no longer applies: GatewayServer.startSocket() already uses SO_REUSEADDR, so bind() succeeds immediately even over TIME_WAIT. Backwards compatibility preserved — callers that need the old behavior can pass startup_sleep=0.25 explicitly.

Provenance

These changes originated as three iterative PRs on a personal fork (byteoak#12, #13, #16), bundled here as one logical "make cold start measurable" surface for upstream review. Full 56-cell CI matrix (Python × Java × OS) green on each byteoak iteration.

Test plan

• All 10 macro scenarios + cold-start scenario pass locally (M-series + JDK 21)
• Existing X1–X8 scenarios unchanged
• python -m py4j.tests.perf --only XA-3,XB,XC --quick passes through the framework runner
• Matrix CI green across Python 3.9–3.13 × Java 8/11/17/21 × ubuntu/macos/windows
• CodSpeed picks up XA-3 / XB / XC / XD as 4 new tracked benchmarks on master after merge