- added xlarge kernel to handle array size >= 2^31
 - ported original pytorch kernel
 - various small fixes

 - remove debug prints
 - use clear macro for 2^31-1 instead of hex

 - fixed gpu indexing error in small kernel
 - make RadixType at least 32 bits

 - also fix axis check error in gpu code

because __props__ should be read only

 - make tests faster
 - remove debug prints

 - GpuTopKOp: uncomment c_code_cache_version
 - add idx_dtype to TopKOp docs

 - add float16 suppport
 - fix 64 bit input bug
 - fix int8 bug due to integer overflow
 - more checking on k
 - add license from pytorch

 - fix thread count calculation bug
 - let write_value and write_index become C macro

 - add multidim support for top_k
 - use unified TopKOp, can implement topk, argtopk, and both

Attempt to lock PR tests to the specific version of libgpuarray that they require.

Add covariance matrix function theano.tensor.cov