xe, benchdnn: fix large buffer ip and add nighly gpu tests

[rjoursler]

Adds large buffer IP tests.

The main IP implementations rely on GEMM and convolution, so correctness is largely handled by previous work. The remaining work in this PR is to the offset arithmetic in the reference convolution and to remove an unuse implementation.

Sample run summary on PVC (the one failure is due to a runtime bug):

tests:26 passed:25 skipped:0 mistrusted:0 unimplemented:0 invalid_arguments:0 failed:1 listed:0
total: 514.01s; fill: 152.07s (30%); compute_ref: 116.66s (23%); compare: 165.29s (32%);

[rjoursler]

make test
disable device_cpu
disable benchdnn_all
enable benchdnn_ip
set test_scope=nightly

[rjoursler]

make test
disable device_cpu
disable benchdnn_all
enable benchdnn_ip