HASTE produces wrong gradients on K80 device

[lmnt-com]

HASTE produces wrong gradients on K80 device

[sharvil]

I double-checked the compute capability of the K80 and it's actually 3.7, not >= 6.0. Since only atomicAdd<double> needed the higher compute capability, I've implemented it in terms of compare-and-swap and lowered the requirements so you can use it on the K80. Could you please try it out and let me know if it works for you?

[amurashov]

Confirmed working! //may be a a good idea to update the docs?

PS Install still required some hacking with Makefile paths to compilers, cuda-libraries, etc. I am fine with this, but some potential users might get turned off by this. // may be a good idea to make a note in the docs that Makefile might require fine-tuning to get up and running? On my systems (including the standard AWS with out-of-the-box TF envs) some voodoo magic was needed for this to compile.