Alloc pickling of GpuArray object

Fix problems in tests in DebugMode.

Faster optimisation

Fix the thunk differences in DebugMode

Fix args with -Wl to nvcc

profile_memory, print the expected peak memory usage if excluding inplace opt.

Enable opt in NanGuardMode. Make it use the default optimizer.

Speed up opt. Now constant folding is done in the eq, no need to traverse the graph to get constant at each node.

fixes #3556: more robust determination of python library name 

add missing return to GpuFromHost R_op

Fix a test on GPU with mode=FAST_COMPILE

Mostly tests fix for 32bit python

Renamed BatchedDotOp to GpuBatchedDot

Fix typo in docs

[WIP] add bessel0

Fix doc gen