[ROCm] Run Sol-Attn on gfx1201 (RDNA4) through aiter

[Shan2L]

Draft: stacked on #1598 (only the last commit, feat(sol_attn): ..., is new here). Depends on aiter
gfx1201_sol_attention (aiter gfx1201 series parts 3-4, upstream PRs follow ROCm/aiter#6390; previews
Shan2L/aiter#2 and Shan2L/aiter#3). This PR will be rebased and marked ready after those are merged.

This PR runs Sol-Attn on AMD Radeon RDNA4 (gfx1201: Radeon AI PRO R9700 / R9600D, RX 9070 series) through aiter.

Today sol_attn needs the CUDA NVlabs sol_attn package; on ROCm SolAttnWeight reports the call as ineligible and
falls back to SDPA (or raises with strict).

Changes

• lightx2v/common/ops/attn/sol_attn.py, ROCm only (CUDA path unchanged):
  • the kernel is aiter.gfx1201_sol_attention, which takes the same arguments as the NVlabs sol_attn forward
    (tau, diag threshold, sink tokens, softmax scale); kv_splits does not apply to it and is ignored;
  • eligibility: gfx1201 only; other ROCm GPUs get an ineligibility reason (strict raises, otherwise the existing
    SDPA fallback);
  • set_config raises on ROCm for thresh_type other than diag and compile_mode other than default;
  • aiter_sage_attn ([ROCm] Add aiter_sage_attn attention backend for gfx1201 (RDNA4) #1598) is accepted as dense_backend;
  • the "Sol-Attn active" log prints the GPU arch on ROCm.
• configs/platforms/amd_rocm/minimax_h3_ref2av_r9700_sol.json: the [ROCm] Add aiter_sage_attn attention backend for gfx1201 (RDNA4) #1598 MiniMax-H3 config with attn_type: sol_attn
  (tau 1.5, diag, dense_steps 6, dense_layers [0], dense backend aiter_sage_attn, strict).
• scripts/platforms/amd_rocm/run_minimax_h3_ref2av_r9700.sh: sol option.

sol_attn is selected only by attn_type; no defaults change.

Test

2 x R9700 (gfx1201, 32 GB), ROCm 7.1, PLATFORM=amd_rocm, MiniMax-H3 ref2av TP2, input street_dance
(girl.png, img_0.jpg), seed 42, 768x1344, 362 frames, DiT block offload, AdaLN cache. Only the DiT self-attention
differs between runs.

• Unit: SolAttnWeight output (3D/4D, softmax scale) and the dense guard (by step, by layer) bitwise identical to direct
  aiter.gfx1201_sol_attention / aiter.gfx1201_sage_attention calls; unsupported thresh_type / compile_mode and
  FP32 inputs with strict raise.
• ruff check / ruff format --check (v0.11.0) clean.

End to end on main (4fe984c + #1595 + #1596, 3 steps, dense_steps 1, ROCM_DETERMINISTIC_FP32_BLAS=1):
this PR vs main without it, with the same aiter ops installed as the sol_attn / sageattention packages (stock
CUDA code path of SolAttnWeight, sage_attn2 dense guard):

	DiT s/step (step 1 dense guard, steps 2-3 Sol)	pipeline total
this PR (dense_backend aiter_sage_attn)	89.3 / 54.2 / 54.7	513.9 s
main + packages	93.1 / 53.9 / 54.4	511.8 s

Video and audio bitwise identical. Per rank 98 Sol calls and 52 dense-guard calls. For comparison, dense
aiter_sage_attn on main (#1598, other GPU pair) runs at 89.0 s/step, so a Sol step is 1.63x faster.

End to end at the stock 29 steps (LightX2V 0.5.0 unmodified, sol_attn with the sol_attn_setting of
minimax_h3_sol_block_offload.json (the settings of the config added here), the aiter ops installed as the sol_attn /
sageattention packages, i.e. the path this PR is bitwise equal to on main; dense guard on steps 1-6 and layer 0):

attention	DiT s/step (steps 2-6 / 7-29 / mean 2-29)	pipeline total	vs BF16 SDPA (s/step / total)
BF16 SDPA	211.4 / 211.7 / 211.6	6450 s
dense gfx1201_sage_attention (code object)	89.0 / 89.1 / 89.1	2877 s	2.38x / 2.24x
Sol, gfx1201_sol_attention HIP kernel	89.4 / 58.1 / 63.7	2144 s	3.32x / 3.01x
Sol, gfx1201_sol_attention code object	89.5 / 54.5 / 60.8	2091 s	3.48x / 3.08x

Per rank 1127 Sol calls and 323 dense-guard calls, 0 fallbacks, all outputs finite.

decoded output pair	video PSNR (min frame)	SSIM (min frame)	audio SNR / cos
Sol code object vs BF16 SDPA	12.80 dB (10.91)	0.522 (0.408)	1.51 dB / 0.661
Sol HIP kernel vs BF16 SDPA	12.82 dB (10.88)	0.523 (0.408)	1.48 dB / 0.660
dense code object vs BF16 SDPA	13.59 dB (10.93)	0.578 (0.416)	6.13 dB / 0.877
BF16 SDPA K/V reversed vs BF16 SDPA (noise floor)	14.64 dB (10.30)	0.629 (0.391)	9.16 dB / 0.938

Over 29 steps the sampling trajectory is chaotic (reordering the FP32 sums of exact attention alone gives 14.6 dB
against the SDPA run), so pixel metrics against one reference run do not measure attention accuracy. Global statistics
of the Sol run (luma 70.0 / 56.9, sharpness 40.2, audio RMS 0.249) are close to the
dense run (69.3 / 56.9, 40.6, 0.238). Sol is block-sparse and lossy by design; its accuracy against FP32 attention
on real MiniMax-H3 tensors is reported with the aiter op.

29-step frames 0/120/241/361 (columns: BF16 SDPA, dense code object, Sol HIP kernel, Sol code object):

Before ready for review

• aiter gfx1201_sol_attention merged
• [ROCm] Add aiter_sage_attn attention backend for gfx1201 (RDNA4) #1598 merged, rebase
• Run scripts/platforms/amd_rocm/run_minimax_h3_ref2av_r9700.sh sol as shipped (29 steps, with the AdaLN cache
  step) on the merged aiter and post the numbers here