RoPE kernel

[ethche]

Adds Helion RoPE kernel along with tuned configs for tritonbench stock shapes. The kernel is structured as a fused Q/K RoPE pass over (batch, seq) tiles. Each program:

• loads cos/sin once for the tile
• uses hl.split / hl.join to operate on the two rotary halves
• applies RoPE to both Q and K in the same kernel
• writes the rotated outputs directly, avoiding separate per-tensor launches
• pre-tuned configs saved in pretuned_kernels/rope

This PR makes some additional changes to test_pretuned_kernels.py. The main change is to re-structure the performance test to enable new hardware other than B200.

Summary results:

device  type  Liger avg speedup  Inductor avg speedup  Helion avg speedup  Helion accuracy
------  ----  -----------------  --------------------  ------------------  ---------------
H100    fwd               2.73x                 4.61x               5.60x                1
H100    bwd               2.37x                 4.91x               5.63x                1

All results:

## Forward, H100

(H, T)          baseline ms       Liger ms  Liger x   Inductor ms  Inductor x   Helion ms  Helion x  Helion acc
-------------  -----------  -------------  -------  ------------  ----------  ----------  --------  ----------
(8192, 1024)      0.230880       0.085056    2.71x      0.046240       4.99x    0.040512     5.70x           1
(8192, 2048)      0.450176       0.169280    2.66x      0.085408       5.27x    0.084864     5.30x           1
(8192, 4096)      0.899168       0.342080    2.63x      0.183712       4.89x    0.166528     5.40x           1
(8192, 8192)      1.778464       0.674784    2.64x      0.485568       3.66x    0.323904     5.49x           1
(8192, 16384)     3.898144       1.331168    2.93x      1.180832       3.30x    0.635168     6.14x           1
(512, 2048)       0.051968       0.018784    2.77x      0.010272       5.06x    0.008320     6.25x           1
(2048, 2048)      0.117920       0.041536    2.84x      0.026592       4.43x    0.021216     5.56x           1
average           0.984600       0.353972    2.73x      0.263068       4.61x    0.170636     5.64x           1

## Backward, H100

(H, T)          baseline ms       Liger ms  Liger x   Inductor ms  Inductor x   Helion ms  Helion x  Helion acc
-------------  -----------  -------------  -------  ------------  ----------  ----------  --------  ----------
(8192, 1024)      0.302944       0.129216    2.34x      0.073280       4.13x    0.072768     4.16x           1
(8192, 2048)      0.625056       0.266400    2.35x      0.085824       7.28x    0.085248     7.33x           1
(8192, 4096)      1.245312       0.526784    2.36x      0.164448       7.57x    0.166496     7.48x           1
(8192, 8192)      2.480416       1.029600    2.41x      0.444544       5.58x    0.323904     7.66x           1
(8192, 16384)     5.275840       2.034432    2.59x      1.181600       4.47x    0.634048     8.32x           1
(512, 2048)       0.151616       0.060800    2.49x      0.069152       2.19x    0.095552     1.59x           1
(2048, 2048)      0.256672       0.109024    2.35x      0.105120       2.44x    0.053088     4.83x           1
average           1.370692       0.552772    2.41x      0.276224       5.12x    0.189544     6.09x           1

[choijon5]

@ethche could you pretune this for B200 also?