Regenerate test_conv_model.cc

[veblush]

A recent QEMU upgrade appears to expose an underlying unaligned memory access issue on Cortex-M3. This is causing crashes (uncaught target signal 7 (Bus error)) in two tests, TestRecordsPersistentTfLiteTensorData and TestRecordsPersistentTfLiteTensorQuantizationData, which were temporarily disabled in PR #3206.

The crash occurs when accessing the test model's Flatbuffer memory, as shown by the following callstack:

#0  0x00008e14 in flatbuffers::Vector<long long, unsigned long>::Get (this=<optimized out>, i=110405) at tensorflow/lite/micro/tools/make/downloads/flatbuffers/include/flatbuffers/vector.h:178
#1  0x00008edc in tflite::internal::InitializeTfLiteTensorFromFlatbuffer (persistent_buffer_allocator=0x407fdc80, non_persistent_buffer_allocator=0x407fdc7c, allocate_temp=false, flatbuffer_tensor=..., 
    buffers=0x15d29 <kTestConvModelData+116>, result=0x407fdbb4) at tensorflow/lite/micro/micro_allocator.cc:285
#2  0x000093c8 in tflite::MicroAllocator::PopulateTfLiteTensorFromFlatbuffer (this=this@entry=0x407fdbd4, model=model@entry=0x15cd9 <kTestConvModelData+36>, tensor=tensor@entry=0x407fdbb4, tensor_index=tensor_index@entry=0, 
    subgraph_idx=<optimized out>, subgraph_idx@entry=0, allocate_temp=<optimized out>, allocate_temp@entry=false) at tensorflow/lite/micro/micro_allocator.cc:1110
#3  0x000098ac in tflite::RecordingMicroAllocator::PopulateTfLiteTensorFromFlatbuffer (this=0x407fdbd4, model=0x15cd9 <kTestConvModelData+36>, tensor=0x407fdbb4, tensor_index=0, subgraph_index=0, allocate_temp=false)
    at tensorflow/lite/micro/recording_micro_allocator.cc:236
#4  0x0000910e in tflite::MicroAllocator::AllocatePersistentTfLiteTensor (this=0x407fdbd4, model=0x15cd9 <kTestConvModelData+36>, subgraph_allocations=0x0, tensor_index=0, subgraph_index=0) at tensorflow/lite/micro/micro_allocator.cc:814
#5  0x00008484 in main (argc=<optimized out>, argv=<optimized out>) at tensorflow/lite/micro/recording_micro_allocator_test.cc:169

My hypothesis is that the pre-generated test model file has an alignment defect although I failed to find why it was shaped that way. Regenerating the model resolves the crash, which supports this theory. However, the new model breaks other tests that depend on specific hardcoded values from the original model.

Since the test model generator does not guarantee this (as it just uses the converter which doesn't promise it), the best path forward is to update the affected tests to be more robust and remove their dependency on specific model data.

[ddavis-2015]

@veblush All the required corrections are in PR #3197