Fix Program::GetIR to handle programs with multiple devices

[9prady9]

Fixes #391

Local tests on a system with multiple NVIDIA GPUs passed. Regular run on a single GPU system also looks good.

[CNugteren]

Thanks a lot, good to hear it works. Code looks good as well. Could you add two minor things to the PR and then I'll merge it:

1. Add a line in the file CHANGELOG with something like - Fixed a bug in the caching when using a context with multiple devices
2. Add a comment in the function you modified at the loop over bin-sizes indicating that we make the assumption here that there is only one valid program (which is valid for the current use of GetIR).

[9prady9]

Will do, I shall send in the changes by tomorrow EOD.