Fix Linear layer bias gradient computation; add size checks to CUDA functions

[hweom]

What does this PR accomplish?

• 🩹 Bug Fix

Closes #169

Changes proposed by this PR:

1. Add an assert to CUDA copy() function to check that source and destination have the same size.
2. Add an assert to CUDA gemm() function to check that computed matrix multiplication dimensions match the passed tensor sizes.
3. Fix the Linear layer bias gradient computation by summing all gradients in the batch instead of copying the output gradient to the bias gradient (which is incorrect since the output gradient contains N items, not 1).

After this change, cuda-memcheck no longer finds errors.

Notes to reviewer:

All unit tests continue to pass except ui which is also broken at HEAD.

📜 Checklist

• [x] The juice-examples run just fine

[hweom]

Test fixes are in #172.

[drahnr]

Could you rebase this PR onto master? I'll merge right away then

[hweom]

Done.