Test each individual driver in random graph tests.

* Use introspection and control API at NDK level to choose each individual
  device.

* For device support API level >= Q (29), we enforce the check on return
  status and buffer value accuracy. For pre-Q devices, we only need them
  to not crash.

* Define accuracy criteria on buffer values. Define 3 sloppy levels for
  single-operation graph and two additional criteria on small and large
  graph test.

* For debugging purpose, automatically dump the spec file when test fails.

Bug: 129569810
Test: NeuralNetworksTest_static_fuzzing
Change-Id: Id592a73e2b74479a9b03c6231d2d14e09dee6b7f