[KernelGen][Nvidia] Add rot90 operator with Triton kernel

[XDYuanzhuLee]

Summary

Adds a Triton kernel for rot90. Triton kernel implementation for rot90.

Testing

• Validated against reference on device via to_reference(inp, True)
• Tested on: Nvidia, Tianshu, Muxi, Ascend, Hygon

Performance

Test command: pytest benchmark/test_rot90.py --level core (NVIDIA H20)

dtype	Size	k	dims	Torch Latency (ms)	Gems Latency (ms)	Speedup
float16	[64, 64]	1	[0, 1]	0.011264	0.007168	1.571
float16	[128, 128]	1	[0, 1]	0.010240	0.008192	1.250
float16	[256, 256]	1	[0, 1]	0.010240	0.009216	1.111
float16	[512, 512]	1	[0, 1]	0.010240	0.010240	1.000
float16	[1024, 1024]	1	[0, 1]	0.014336	0.017408	0.824
float16	[2048, 2048]	1	[0, 1]	0.026624	0.046080	0.578
float16	[100, 200]	1	[0, 1]	0.010240	0.008192	1.250
float16	[200, 400]	1	[0, 1]	0.010240	0.008192	1.250
float16	[400, 800]	1	[0, 1]	0.011264	0.011264	1.000
float32	[64, 64]	1	[0, 1]	0.009216	0.008192	1.125
float32	[128, 128]	1	[0, 1]	0.010240	0.008192	1.250
float32	[256, 256]	1	[0, 1]	0.010240	0.008192	1.250
float32	[512, 512]	1	[0, 1]	0.011264	0.011264	1.000
float32	[1024, 1024]	1	[0, 1]	0.016384	0.020480	0.800
float32	[2048, 2048]	1	[0, 1]	0.036864	0.058368	0.632
float32	[100, 200]	1	[0, 1]	0.010240	0.008192	1.250
float32	[200, 400]	1	[0, 1]	0.010240	0.009216	1.111
float32	[400, 800]	1	[0, 1]	0.012288	0.012288	1.000
bfloat16	[64, 64]	1	[0, 1]	0.009216	0.007168	1.286
bfloat16	[128, 128]	1	[0, 1]	0.009216	0.008192	1.125
bfloat16	[256, 256]	1	[0, 1]	0.010240	0.009216	1.111
bfloat16	[512, 512]	1	[0, 1]	0.011264	0.010240	1.100
bfloat16	[1024, 1024]	1	[0, 1]	0.015360	0.017408	0.882
bfloat16	[2048, 2048]	1	[0, 1]	0.027648	0.045056	0.614
bfloat16	[100, 200]	1	[0, 1]	0.010240	0.008192	1.250
bfloat16	[200, 400]	1	[0, 1]	0.010240	0.009216	1.111
bfloat16	[400, 800]	1	[0, 1]	0.011264	0.011264	1.000
Arithmetic Mean	—	—	—	—	—	1.075

Multi-backend Testing

Backend	Accuracy Test	Speedup (mean)	Notes
Nvidia (H20)	PASS (27 cases)	1.075	Primary
Tianshu	PASS	1.242	—
Muxi	PASS	1.151	—
Ascend	PASS	—	—
Hygon	FAIL	0.911	error: operand #0 does not dominate this use

Files Changed

• src/flag_gems/ops/rot90.py: Triton kernel implementation
• tests/test_rot90.py: Accuracy test
• benchmark/test_rot90.py: Performance benchmark
• src/flag_gems/ops/__init__.py: Register import and __all__
• src/flag_gems/__init__.py: Register to _FULL_CONFIG
• conf/operators.yaml: Add operator entry (kind: Math, stage: alpha 5.1)