[LIBCLC][AMDGCN] Fix get_max_sub_group_size

[npmiller]

Using defines to figure out the wavefront size there is incorrect
because libclc is not built for a specific amdgcn version, so it will
always default to 64.

Instead use the __oclc_wavefront64 global variable provided by ROCm,
which will be set to a different value depending on the architecture.

This may fix some of the discrepancies between tests being run on gfx908 and the tests running on the CI, as the CI hardware uses a wavefront of 32 which mismatches with what was returned by this function, and that this function is used in the implementation of the group collectives.

And so it may fix:

• [SYCL] Mark broadcast tests XFAIL for HIP. llvm-test-suite#740
  and potentially the discrepancies I was seeing on:
• [SYCL][AMDGCN] Fix up and down shuffles and reductions #5359

[npmiller]

So the test results:

********************
Failed Tests (1):
  SYCL :: GroupAlgorithm/SYCL2020/shift_left_right.cpp

********************
Unexpectedly Passed Tests (3):
  SYCL :: SubGroup/broadcast.cpp
  SYCL :: SubGroup/broadcast_fp16.cpp
  SYCL :: SubGroup/broadcast_fp64.cpp


Testing Time: 190.11s
  Unsupported        : 400
  Passed             : 288
  Expectedly Failed  :  53
  Failed             :   1
  Unexpectedly Passed:   3

So it looks like this does fix the broadcast tests as I was hoping.

I'm not sure what's going on with the new failure, however the test is using shuffleUp and shuffleDown so it's possible that #5359 would fix it. I'm not able to make it fail on gfx908 so it is difficult to debug, maybe I should close this PR and bring the patch into the shuffles PR to see how the testing goes then.

[npmiller]

Closing, will be merged as part of #5359