Fix NNAPI delegation: no scratch tensors and scale 0 for floats

  - TFLite has temporary (scratch) tensors needed for its own CPU
  implementation
  - Delegating temporary tensors to NNAPI will cause validation
  failure, as they are not used
  - This CL skips such temporaries for CONV_2D operations - others may
  need to be added later
  - Setting non-zero scale for 32-bit float tensors causes validation
  errors
  - Also added delegation for MUL.

Change-Id: I62dfaa53f4cdffd0d1dcedb28d181b8c92755ca8
Test: build and run NNAPI benchmarking app