OpenCL BE: set CHIP_USE_INTEL_USM on by default

[linehill]

PVC platforms seems to benefit a lot from the USM. Few picks:

• HeCBench/mandelbrot: 56x speed-up.

• HeCBench/floydwarshall: 14x speed-up.

• HeCBench/bitonic-sort: 33x speed-up.

• HeCBench/nw: 20x speed-up.

• HeCBench/fft: 11x speed-up.

• And bunch of other HeCBench cases with speed-ups between 1.16x and 6.5x.

The OpenCL backend still falls back to use SVM (the previous default) when the USM is not available.

[pvelesko]

@linehill still failing hipKernelLaunchIsNonBlocking

[linehill]

Added work-around to a test timeout issue.