It does GEMM on two BF16 matrices and Accumulates in F32 matrix

Source oneapi before configuring IMEX.
Remove redundant SYCL_DIR.

Assumptions:

Github actions is serialized. (Running one action at a time)
CI uses system installed cmake (not from anaconda)

List of changes:

Remove usage of anaconda
Does not cache LLVM build
Local persistent LLVM source repo is shared between cpu and gpu CI. We can do this since CI actions are serialized. cpu and gpu CI actions will pull upstream change an checkout LLVM commit id based on PR.
LLVM/MLIR is always built from scratch
IMEX is built along LLVM as an external project instead of a separate build. This setup cuts LLVM build time since it reduce the number of targets need to be built for LLVM.
Single cmake target "check-imex" is invoked. This target both builds IMEX and runs test.