Split tensor arena into persistent and non-persistent arenas

[andresovela]

This PR is an attempt to fix xmos/ai_tools#908

I renamed the SHARED_TENSOR_ARENA definition to EXTERN_TENSOR_ARENA, and I added a new SPLIT_PERSISTENT_TENSOR_ARENA definition.

This PR adds the following when SPLIT_PERSISTENT_TENSOR_ARENA is defined:

• A new extern uint8_t persistent_tensor_arena[]; declaration
• Code uses this new buffer only when defined

I had initially renamed tensor_arena to non_persistent_tensor_arena, but I found that I had to generate the huge tflTensors[] array twice just because of the new name, so I decided against it. The code would have looked like this otherwise:

#ifdef SPLIT_PERSISTENT_TENSOR_ARENA
TfLiteTensor tflTensors[] = 
{{ {(int32_t*)(non_persistent_tensor_arena + 26112)},(TfLiteIntArray*)&g0.tensor_dimension0, kTfLiteInt8, {kTfLiteAffineQuantization, const_cast<void*>(static_cast<const void*>(&g0.quant0)) }, {g0.quant0.scale->data[0], g0.quant0.zero_point->data[0] },64, kTfLiteArenaRw, false, },
.
.
.
};
#else
TfLiteTensor tflTensors[] = 
{{ {(int32_t*)(tensor_arena + 26112)},(TfLiteIntArray*)&g0.tensor_dimension0, kTfLiteInt8, {kTfLiteAffineQuantization, const_cast<void*>(static_cast<const void*>(&g0.quant0)) }, {g0.quant0.scale->data[0], g0.quant0.zero_point->data[0] },64, kTfLiteArenaRw, false, },
.
.
.
};
#endif

I haven't been able to test if this would work or not, I just wrote the code for now.

[andresovela]

Note that if this gets merged we would need to update the app_flash_two_models_one_arena example

[andresovela]

I was able to test this feature. See this comment for details.