[SYCL] Reject unsuitable old ocloc from Kernel Compile by passing device IPVersions when loading

[cperkinsintel]

For Kernel Compiler Open CL C support the device.ext_can_compile now passes the device down so we can make sure OCLOC supports commands with the -dev IPVersionStr .

[bashbaug]

This change appears to:

1. Add a fallback on Windows that checks for an ocloc DLL in a fixed location, if one isn't found using standard DLL search orders.
2. Performs an ocloc query when loading the ocloc DLL as a test to determine that the ocloc library is "good".

Both of these changes are improvements over the existing code, hence my review approval, but I think we ought to reconsider how we handle this in the future. For example:

• Can we handle the case where the user installs into a non-standard directory and hence the ocloc DLL isn't found at the fixed location?
• When there are multiple devices in the system, are there cases where we may need multiple ocloc libraries? Should the ocloc library handle be tied to the SYCL context (or perhaps the SYCL device, or the SYCL platform) rather than stored in a single static global variable?
• Can we add some encapsulation so this code is easier to understand and maintain in the future?

[aelovikov-intel]

PR title is still wrong.