Add global kernel cache and HCF infrastructure

[illuhad]

• Add common global kernel cache infrastructure
• Add unified HCF file format for kernels, tools, and documentation
• Reimplement explicit multipass for CUDA and SPIR-V to use new infrastructure
• Reimplement syclcc explicit multipass drivers to utilize unified HCF generation and integration header generation

TODO: Programs don't yet terminate cleanly on Level Zero due to a lifetime issue of the Level Zero backend shutting down before the cached SPIR-V kernels are unloaded.

SYCL-Bench task throughput benchmarks indicate that CUDA explicit multipass with the new kernel cache infrastructure outperforms integrated multipass by 30-40%.

[illuhad]

It seems there is no easy fix for the Level Zero issue. This relates to the fundamental issue that we need to transition to a runtime that has its lifetime dependent on the existence of SYCL objects. This is a larger change, and should not block this PR.

[illuhad]

In it goes.