Support zero-sized tensor and check for omitted input in CpuExecutor.

Zero-sized tensors are only supported internally to the driver, or
reported in output shapes. It is illegal to pre-specify a zero-sized
tensor as model input or output.

To summarize the meanings of dimension = 0:
- Dimension = 0 for model input: dynamic input and must be fully-specified
at execution time
- Dimension = 0 for internal operand / model output: unknown, to be
deduced from execution
- Dimension = 0 from getOutputOperandDimensions:
  * If NO_ERROR, it is a zero-sized output
  * If OUTPUT_INSUFFICIENT_SIZE, it is unknown

Add two additional fields in OperationRegistration:
- allowOmittedOperand, if false, CpuExecutor will enforce inputs/outputs
are not null
- allowZeroSizedInput, if false, CpuExecutor will enforce every dimension
of inputs not being 0

The current implementation is assuming that none of the operations with
the old switch statement path support zero-sized input. Only the
operations with OperationResolver can support zero-sized input tensor.

All the operations with OperationResolver are reporting false for
allowZeroSizedInput. Will enable this on a small subset of operations with
separate CLs.

Bug: 126737477
Test: NeuralNetworksTest_static
Change-Id: Ia94d67e4c8c6a49b543d29ebb3b31d509ece0970
Merged-In: Ia94d67e4c8c6a49b543d29ebb3b31d509ece0970
(cherry picked from commit 5750766c)