Combined hash and fill AES loop

[SChernykh]

Adds more parallelizm into AES loop so modern CPUs can take advantage of it. Also, scratchpad data moves between L1 and L3 caches only one time which saves time and energy per hash.

[hyc]

For typechecking purposes you could still typedef this param to uint64_t foo[8] and use the type in the parameter lists.

[tevador]

Yeah, uint64_t tempHash[8] could be more readable in the API and it decays into a pointer anyways.

[hyc]

And that would have prevented that sizeof() bug too.

[c0mm4nd]

Using the "first" and "next" hash function in randomx.cpp, the miner (mine which based on this repo, not xmirg) receives "Low difficulty share" from pool.

[SChernykh]

You have to be careful using these new functions, check how benchmark.cpp handles them and where it updates the nonce.

[c0mm4nd]

My miner works fine on this workflow

	uint64_t tempHash[8];
	while (nonce < noncesCount) {
		nonce = atomicNonce.fetch_add(1);
		store32(noncePtr, nonce);
	        randomx_calculate_hash_first(vm, tempHash, blockTemplate, sizeof(blockTemplate));
		randomx_calculate_hash_next(vm, tempHash, blockTemplate, sizeof(blockTemplate), &hash);
		result.xorWith(hash);
	}

But when the randomx_calculate_hash_first func is out of while loop, the miner gets the wrong result.

	uint64_t tempHash[8];

	store32(noncePtr, nonce);
	randomx_calculate_hash_first(vm, tempHash, blockTemplate, sizeof(blockTemplate));

	while (nonce < noncesCount) {
		nonce = atomicNonce.fetch_add(1);
		store32(noncePtr, nonce);
		randomx_calculate_hash_next(vm, tempHash, blockTemplate, sizeof(blockTemplate), &hash);
		result.xorWith(hash);
	}