[KernelGen][Kunlunxin] Add bernoulli operator

[NineAnnAnn]

Summary

Adds a Kunlunxin (P800) backend specialization for bernoulli. Draws binary random numbers (0 or 1) from a Bernoulli distribution parameterized by the input probability tensor. Adds a dedicated Triton kernel for this backend and registers it in the backend ops module.

Testing

• Validated against reference on device (Kunlunxin (P800))
• Accuracy: PASS (6 cases) over 3 dtypes x 6 shapes

Performance

Test command: pytest benchmark/test_bernoulli.py --level core (Kunlunxin (P800))

bernoulli

dtype	Size	Torch Latency (ms)	Gems Latency (ms)	Speedup
float16	[64, 64]	0.011379	0.006026	1.888
float16	[4096, 4096]	3.374843	0.209165	16.135
float16	[64, 512, 512]	3.374927	0.208959	16.151
float32	[64, 64]	0.011972	0.005828	2.054
float32	[4096, 4096]	3.345646	0.211519	15.817
float32	[64, 512, 512]	3.345941	0.211849	15.794
bfloat16	[64, 64]	0.011650	0.005710	2.040
bfloat16	[4096, 4096]	3.269944	0.315181	10.375
bfloat16	[64, 512, 512]	3.271776	0.315200	10.380

Operator	Geometric Mean Speedup
bernoulli	7.252

Multi-backend Testing

Backend	Accuracy Test	Speedup (mean)	Notes
Kunlunxin (P800)	PASS (6 cases)	7.252	Primary

Files Changed

• src/flag_gems/runtime/backend/_kunlunxin/ops/bernoulli.py: New Kunlunxin (P800) kernel implementation
• src/flag_gems/runtime/backend/_kunlunxin/ops/__init__.py: Register import and __all__ entry

[Dongxu-H]

/test | bernoulli:kunlunxin

[github-actions]

❌ On-demand test failed

Operator: bernoulli
Runner: kunlunxin
Backend: kunlunxin

The test failed to complete. Check the workflow run for details.

[Dongxu-H]

init.py 中 bernoulli 的导入位置应该按字母顺序放在 batch_norm 和 bernoulli_ 之间
all 列表中 bernoulli 应该按字母顺序排列而不是放在末尾