scikit-sparse install problems don't fail meaningfully

[paulthebaker]

@siyuan-chen found an issue in a fresh enterprise install using pip.

scikit-sparse (sks) depends on the SuiteSparse C library. If SuiteSparse isn't installed, sks can still install "successfully", but it will throw errors when used.

In LogLikelihood.__call__() there's a cholesky call inside a try / except that catches all errors. This means that if sks isn't linked to SuiteSparse correctly, the sks error will send the code to the except statement. LogLiklihood will return -inf when it should crash and exit. There's no indication to the user what went wrong.

The install problem can happen because requirements.txt installs sks with the --no-binary option. This assumes the user has an existing install of SuiteSparse. Installing the packaged binary of sks will use a generic BLAS, which is slower than a locally optimized version. If you install packaged binaries of numpy and scipy via pip, they won't used optimized BLAS either.

There are few things to do:

• additionally check sks linking when we check that it is installed and give a meaningful error message
• change the requirements.txt to use the packaged binary version of sks. A pip install is already using generic BLAS for numpy... 🤷
• specify which errors that try / except is intended to catch, so the code exits on others. I think we want it to return -inf and continue if Sigma is singular

[jellis18]

This is somewhat related to #257 and #258 and will somewhat be solved by this PR I could possible add a check in here so the install will fail properly if the linked libraries aren't found.