Enable shared memory by default

[kurapov-peter]

The PR adds proper code generation for Intel GPUs.

Notes:

• The shared buffer is allocated statically within the genx.ll module and its size is not checked for now.
• L0 backend now uses smem by default.
• There was an issue with cgen: emitCall would use the wrong runtime module. This is now fixed by patching the module triple upon extension module load.
• We currently produce some redundant address space casts between generic and shared-local. Those produce actual runtime checks and affect perf. I left the warnings in the log.

[Garra1980]

Any specific tests for sh mem or existing set is enough?

• We currently produce some redundant address space casts between generic and shared-local. Those produce actual runtime checks and affect perf. I left the warnings in the log.

Is it worth to be handled as a separate issue?

[kurapov-peter]

Yes, we'll have some smem specific tests. Address space casts will be addressed separately.

[kurapov-peter]

The pass rate is now decent, so this is ready for review.

[alexbaden]

What happens if shared_mem is turned off? What function implementation do we fall back to?

[kurapov-peter]

Just noticed the comment. I haven't tested it, but what should happen is it should emit the agg_count from the "cpu" runtime module instead. BTW, I'm not sure if we need the option to disable shared memory at all :)