Register activation kernel layers on XPU

[jiqing-feng]

Summary

kernels-community/activation now ships an XPU backend, so map the six layers
it provides to XPU in addition to CUDA. They were previously registered for
CUDA only, which left XPU on the eager fallback.

Affected entries in _KERNEL_MAPPING: FastGELU, QuickGELU, NewGELU,
SiLU, GeLU, GeluTanh. Each reuses the CUDA repository, layer name and
version, under Mode.INFERENCE | Mode.TORCH_COMPILE.

FastGELU, NewGELU and QuickGELU have no native PyTorch operator and
expand into several element-wise launches in eager mode, so the fused kernel is
several times faster. SiLU, GeLU and GeluTanh read and write the same
amount of data as the corresponding PyTorch operator and are on par with it;
they are mapped for parity with CUDA.

Validation

Checked on Intel Arc Pro B60 with torch 2.13.0+xpu. All six layers run in
eager mode and under torch.compile, with compiled and eager results
identical:

Layer	speedup over eager	max diff vs PyTorch reference
FastGELU	6.20x	1.95e-03
NewGELU	5.40x	1.95e-03
QuickGELU	3.44x	1.95e-03
SiLU	0.98x	0.00e+00
GeLU	0.98x	0.00e+00
GeluTanh	1.18x	0.00e+00

Speedups are float16 at 16384 x 8192. The differences for the three
composite activations are inherent to their approximations, which do not match
the PyTorch operator they are compared against, and are the same on CUDA.

Dependencies

Requires huggingface/kernels-community#1071, which adds the XPU backend to
activation, and huggingface/kernels#752, which that backend depends on. Both
must be merged and an activation build carrying the XPU variant published on
the Hub before this mapping resolves.

[github-actions]

---

CI recap

Dashboard: View test results in Grafana
Latest run: 31353982233:2
Result: success | Jobs: 16 | Tests: 166,813 | Failures: 0 | Duration: 15h 59m

[Rocketknight1]

cc @IlyasMoutawwakil

[IlyasMoutawwakil]

@jiqing-feng tell me when it's ready