[NeoML] Layers mem-optimize

[favorart]

The max consumption of temporary memory is reduced by layers:

• CBatchNormalizationLayer.runWhenLearning()

  • from objectSize + inputSize to max( inputSize, objectSize )
  • tested

• CBatchNormalizationLayer.backwardWhenLearning()

  • from outputDiffSize +3* objectSize to outputDiffSize +1* objectSize
  • tested

• CBinaryCrossEntropyLayer.BatchCalculateLossAndGradient()

  • from 9 to 3 (* batchSize)
  • tested

• CCenterLossLayer.BatchCalculateLossAndGradient()

  • from 3 to 2 (* inputDataSize) and 3 to 2 * (classCentersSize)
  • tested

• CFocalLossLayer.BatchCalculateLossAndGradient()

  • from (dataSize +3* batchSize) to (max( dataSize, batchSize ) + batchSize)
  • tested

• CGELULayer.backwardFastApproximate()

  • from 2 to 1 (* dataSize)
  • tested

• CPrecisionRecallLayer .RunOnceAfterReset()

  • from 8 to 3 (* vectorSize)
  • tested