perfetto: cache Android SDK track UUIDs and simplify event checks

Cache the process track UUID after native initialization and derive
thread UUIDs in Java to avoid repeated JNI calls. Preserve the JNI
fallback and signed 32-bit tid conversion.

Remove the redundant isRegistered() check before isEnabled(). Add tests
for UUID compatibility and category re-registration, plus a host
benchmark via tools/run_android_sdk_host_test --benchmark.

Recorded Linux host CPU ns/event: lower median of two rounds of
5 x 200 ms, one JVM per scenario, pinned to four cores. These runs
preceded the final helper removal; no rerun was performed.

ns/event          base UUIDs measured  noise vs base
disabled_instant  18.2  17.9     17.5 0.168%  -4.16%
disabled_args     18.3  17.9     16.7  5.43%  -8.59%

Base precedes this commit; UUIDs adds caching. Measured includes the
event-path optimizations from that run. Noise is the largest difference
between measurement rounds.

Enabled instant, counter and named-track results were within noise.
Unrelated paths shifted by 2-4% between builds. Allocation and trace
bytes per event were unchanged.

Android device performance has not been measured. HotSpot ignores
@FastNative and @CriticalNative, so these results do not establish the
gain on ART.