allocator initial cpu memory usage is too large

[tensor-tang]

This is an issue related with #11185

Even we can use fraction_of_cpu_memory_to_use to work around, but it's still tricky and does not make sense that init memory with max_chunk_size_ which it 3% memory.

BuddyAllocator::PoolSet::iterator BuddyAllocator::RefillPool() :

Paddle/paddle/fluid/memory/detail/buddy_allocator.cc

Lines 189 to 192 in 6d80dd5

	// Allocate a new maximum sized block
	size_t index = 0;
	void* p = system_allocator_->Alloc(&index, max_chunk_size_);

void* BuddyAllocator::Alloc(size_t unaligned_size) :

Paddle/paddle/fluid/memory/detail/buddy_allocator.cc

Lines 67 to 69 in 6d80dd5

	// refill the pool if failure
	if (it == pool_.end()) {
	it = RefillPool();

[Superjomn]

Maybe an absolute memory usage is better for inference.