Updating v1 monero cryptonight tweaks, and less global read/writes

[vtnerd]

This patch has a few updates:

• A branchless version of the first tweak that is posted in the main PR for monero.
• Removes two global reads/writes per thread per iteration in the inner loop. The reads/writes are now "batched" with the existing ones. Hopefully this will remove the 5% slowdown that was reported, although I have not done in any "speed" runs for comparison.
• Nonce flag should have been wrong after the first nonce (per batch).
• Adds the second XOR tweak - which is partially seen on the monero PR. The changes I made recently were to do a 64-bit XOR over bytes [8...15] instead of just [8...11], and make the constant XOR value being used a combination of 64-bits from the input XORed with the remaining 64-bits of the keccak output. This makes it no longer user controllable.

I ran a test harness with some test outputs that I can provide if necessary. People watching this PR should notice the changes to the CPU version as well - its similar to the ones that will be provided to other CPU miners.

[vtnerd]

I also wanted to mention that alternatively, d_input and ctx_state could have been passed to the inner work function, and then the d_tweak1_2 variable could have been computed without allocating and copying memory to an additional 64-bits of global memory. I implemented it this way because the offsets for the "tweak" are fixed, but the offset for the "nonce" is technically not (even though existing software assumes specific offset). This could bring additional compute savings at the cost of maintenance. Its not difficult to adapt the code if this is desired : pass in startNonce, d_input, and d_ctx_state to the inner processing loop and copy the nonce calculation logic and tweak logic from the prepare function.